Labellings for assumption-based and abstract argumentation

Labellings for assumption-based and abstract argumentation

International Journal of Approximate Reasoning 84 (2017) 110–149 Contents lists available at ScienceDirect International Journal of Approximate Reas...

1MB Sizes 0 Downloads 71 Views

International Journal of Approximate Reasoning 84 (2017) 110–149

Contents lists available at ScienceDirect

International Journal of Approximate Reasoning www.elsevier.com/locate/ijar

Labellings for assumption-based and abstract argumentation Claudia Schulz ∗ , Francesca Toni Department of Computing, Imperial College London, London SW7 2AZ, UK

a r t i c l e

i n f o

Article history: Received 11 August 2016 Received in revised form 21 February 2017 Accepted 23 February 2017 Available online 1 March 2017 Keywords: Assumption-based argumentation Labelling semantics Abstract argumentation

a b s t r a c t The semantics of Assumption-Based Argumentation (ABA) frameworks are traditionally characterised as assumption extensions, i.e. sets of accepted assumptions. Assumption labellings are an alternative way to express the semantics of flat ABA frameworks, where one of the labels in, out, or undec is assigned to each assumption. They are beneficial for applications where it is important to distinguish not only between accepted and nonaccepted assumptions, but further divide the non-accepted assumptions into those which are clearly rejected and those which are neither accepted nor rejected and thus undecided. We prove one-to-one correspondences between assumption labellings and extensions for the admissible, grounded, complete, preferred, ideal, semi-stable and stable semantics. We also show how the definition of assumption labellings for flat ABA frameworks can be extended to assumption labellings for any (flat and non-flat) ABA framework, enabling reasoning with a wider range of scenarios. Since flat ABA frameworks are structured instances of Abstract Argumentation (AA) frameworks, we furthermore investigate the relation between assumption labellings for flat ABA frameworks and argument labellings for AA frameworks. Building upon prior work on complete assumption and argument labellings, we prove one-to-one correspondences between grounded, preferred, ideal, and stable assumption and argument labellings, and a one-to-many correspondence between admissible assumption and argument labellings. Inspired by the notion of admissible assumption labellings we introduce committed admissible argument labellings for AA frameworks, which correspond more closely to admissible assumption labellings of ABA frameworks than admissible argument labellings do. © 2017 Elsevier Inc. All rights reserved.

1. Introduction Argumentation provides an intuitive way of modelling human reasoning and decision making and has thus received considerable attention in AI research [1,2]. Argumentation frameworks formalise the notions of arguments and conflicts between them, and specify semantics to determine which arguments are accepted in a debate of conflicting arguments. Two main types of argumentation frameworks can be distinguished: In Abstract Argumentation (AA) frameworks [3] a set of abstract entities, called arguments, is given along with an attack relation between these arguments. In structured argumentation frameworks (e.g. [4–7], see [8] for an overview) structured knowledge is given from which arguments are constructed and attacks are derived. Here, we deal with a certain type of structured argumentation frameworks, namely Assumption-Based Argumentation (ABA) frameworks [4,9,10], which have proven useful in a variety of applications including agent negotiation and dialogue

*

Corresponding author. E-mail addresses: [email protected] (C. Schulz), [email protected] (F. Toni).

http://dx.doi.org/10.1016/j.ijar.2017.02.005 0888-613X/© 2017 Elsevier Inc. All rights reserved.

C. Schulz, F. Toni / International Journal of Approximate Reasoning 84 (2017) 110–149

111

[11–16], decision making [17–19], web reasoning [20], and explanation [21,22]. In an ABA framework structured knowledge is given in terms of inference rules expressed in an underlying logical language. A subset of sentences is defined to be the set of assumptions, each of which has a contrary sentence in the language. As an example, consider an ABA framework with the following inference rules1 about the excellence of Imperial College London (ICL) and the withdrawal of the UK from the EU, where assumptions are indicated in italic: ICL is an excellent university ← ICL takes many international students and staff Fewer EU citizens apply to ICL ← EU citizens need a visa to move to the UK EU citizens can move to the UK without a visa ← The UK remains in the EU The contraries of the assumptions are as follows:

• contrary of ICL takes many international students and staff : Fewer EU citizens apply to ICL; • contrary of EU citizens need a visa to move to the UK: EU citizens can move to the UK without a visa; • contrary of The UK remains in the EU: The UK leaves the EU. The semantics of ABA frameworks are defined in terms of sets of accepted assumptions, called assumption extensions, which are determined based on the contraries of assumptions and derivations from assumptions using the inference rules [4]. In the above example, one of the assumption extensions is {The UK remains in the EU }, expressing that the assumption The UK remains in the EU is accepted and all other assumptions are not. Given a flat ABA framework, where assumptions do not occur on the left-hand side of inference rules, arguments and attacks between them can be constructed. A flat ABA framework can thus instantiate an AA framework comprising all arguments and attacks constructable from the flat ABA framework [23]. It follows that the semantics of a flat ABA framework can also be expressed in terms of the semantics of AA frameworks [23], i.e. as sets of accepted arguments called argument extensions [3]. Importantly, the assumption and argument extensions of a flat ABA framework correspond for nearly all semantics defined for flat ABA frameworks2 : an argument extension contains all arguments supported by the assumptions in an assumption extension, and conversely an assumption extension contains all assumptions supporting arguments in an argument extension [23,20,24]. The semantics of AA frameworks can be alternatively formulated as argument labellings, which assign one of the labels in (accepted), out (rejected), or undec (undecided) to each argument [25,26]. This notion of semantics has the advantage over argument extensions that it does not only distinguish between accepted and non-accepted arguments, but further divides the non-accepted arguments into rejected and undecided ones. Since argument labellings and argument extensions correspond [25,26], argument labellings can also be used to characterise the semantics for flat ABA frameworks in terms of arguments. Recently, the idea of argument labellings was transferred to assumptions [27], yielding a new characterisation of the semantics of flat ABA frameworks. In contrast to argument labellings which label whole arguments, assumption labellings label each assumption as in (accepted), out (rejected), or undec (undecided). Assumption labellings have the advantage over assumption extensions that rejected (out) assumptions and assumptions which are neither accepted nor rejected (undec) are distinguished. This distinction can be important in applications such as decision making: undecided assumptions can for example provide an indication that further information from an expert is required in order to decide their acceptability for sure. For instance, in the ICL example one of the assumption labellings labels The UK remains in the EU as in, EU citizens need a visa to move to the UK as out, and ICL takes many international students and staff as undec. This expresses that it is not certain whether or not ICL takes many international students and staff and thus provides a more detailed interpretation than the previously given assumption extension. This work considerably extends [27,28], where we considered only two semantics of flat ABA frameworks.3 Here, we investigate all semantics defined for flat ABA frameworks, i.e. admissible, grounded, complete, preferred, ideal, semi-stable, and stable semantics, making this paper a self-contained reference for assumption labellings and their relation with both assumption extensions and argument labellings. We prove that there is a one-to-one correspondence between assumption labellings and extensions for all aforementioned semantics. We also investigate the relation between assumption and argument labellings for flat ABA frameworks, showing a one-to-one correspondence for the grounded, complete, preferred, ideal, and stable semantics, as summarised in Fig. 1. Since semi-stable argument and assumption extensions do not correspond [24], it is unsurprising that the respective labellings do not correspond either, as shown in Fig. 2. Concerning the admissible semantics we prove a one-to-many correspondence between assumption and argument labellings. Based on this dissimilarity, we introduce a variant of admissible argument labellings for AA frameworks, called committed admissible argument labellings, which correspond more closely to admissible assumption labellings than the original admissible argument

1 Rules are adapted from www.imperial.ac.uk/newsandeventspggrp/imperialcollege/newssummary/news_13-1-2016-18-10-57 and www.independent. co.uk/student/news/eu-referendum-result-brexit-leave-remain-higher-education-sector-students-a7100106.html. 2 The only exception is the semi-stable semantics [24]. 3 Caminada and Schulz [29] introduce grounded, preferred, ideal, and stable assumption labellings for ABA but do neither prove correspondence with the respective assumption extensions nor investigate their relation with argument labellings.

112

C. Schulz, F. Toni / International Journal of Approximate Reasoning 84 (2017) 110–149

Fig. 1. A summary of results concerning the complete, grounded, preferred, stable, and ideal semantics in the different sections of this paper, where applicable, in the context of previous work. Bidirectional arrows indicate semantic correspondence, and bold indicates novel work in this paper.

Fig. 2. Results concerning the semi-stable semantics in the different sections of this paper, where applicable, in the context of previous work. Bidirectional arrows indicate semantic correspondence, crossed out arrows denote non-correspondence, and bold indicates novel work in this paper.

labellings, as illustrated in Fig. 3. We furthermore extend [27,28], where only flat ABA frameworks have been considered, by introducing assumption labellings for any ABA framework. Our results are summarised in Fig. 4. In ABA frameworks which may not be flat, assumptions can occur on the left-hand side of an inference rule and can thus constitute facts, expressing statements such as “I firmly believe that ICL will always be an excellent university”. The paper is organised as follows. In Section 2 we give background on flat ABA frameworks and AA frameworks. In Section 3 we introduce assumption labellings for the different semantics of flat ABA frameworks and prove their correspondence with assumption extensions of flat ABA frameworks, building upon prior work in [27,28]. In Section 4 we simplify the definition of assumption labellings for flat ABA frameworks by considering only certain sets of assumptions as attackers of assumptions. We furthermore introduce a graphical representation of flat ABA frameworks and illustrate how assumption labellings can be easily determined and represented using these graphs. In Section 5 we investigate the correspondence between assumption and argument labellings of flat ABA frameworks (considerably extending preliminary work in [27]) and introduce committed admissible argument labellings as a variant of admissible argument labellings for AA frameworks. In Section 6 we extend the definition of assumption labellings from flat ABA frameworks to any ABA framework, and in Section 7 we conclude and discuss future research. 2. Background 2.1. Abstract argumentation An Abstract Argumentation (AA) framework [3] is a pair  Ar , Att , where Ar is a set of arguments and Att ⊆ Ar × Ar is a binary attack relation between arguments. A pair ( A , B ) ∈ Att expresses that argument A attacks argument B, or equivalently that B is attacked by A. A set of arguments Args ⊆ Ar attacks an argument B ∈ Ar if and only if there is A ∈ Args such that A attacks B. Args+ = { A ∈ Ar | Args attacks A } denotes the set of all arguments attacked by Args [26].

C. Schulz, F. Toni / International Journal of Approximate Reasoning 84 (2017) 110–149

113

Fig. 3. Results concerning the admissible semantics in the different sections of this paper, where applicable, in the context of previous work. Bidirectional arrows indicate semantic correspondence (arrows with the same starting point but different end points indicate one-to-many correspondence), and bold indicates novel work in this paper.

Fig. 4. Results for any (possibly non-flat) ABA framework in the different sections of this paper, where applicable, in the context of previous work. Bidirectional arrows indicate semantic correspondence, and bold indicates novel work in this paper.

Let Args ⊆ Ar be a set of arguments.

• Args is conflict-free if and only if Args ∩ Args+ = ∅. • Args defends A ∈ Ar if and only if Args attacks every B ∈ Ar attacking A. The semantics of an AA framework are defined in terms of argument extensions, i.e. sets of accepted arguments [3,23,30]. A set of arguments Args ⊆ Ar is

• • • •

an admissible argument extension if and only if Args is conflict-free and defends all arguments A ∈ Args; a complete argument extension if and only if Args is conflict-free and consists of all arguments it defends; a grounded argument extension if and only if Args is a minimal (w.r.t. ⊆) complete argument extension; a preferred argument extension if and only if Args is a maximal (w.r.t. ⊆) complete argument extension;

114

C. Schulz, F. Toni / International Journal of Approximate Reasoning 84 (2017) 110–149

• an ideal argument extension if and only if Args is a maximal (w.r.t. ⊆) admissible argument extension satisfying that for all preferred argument extensions Args , Args ⊆ Args ; • a semi-stable argument extension if and only if Args is a complete argument extension and for all complete argument extensions Args , Args ∪ Args+ ⊂ Args ∪ Args + ; • a stable argument extension if and only if Args is a complete argument extension and Args ∪ Args+ = Ar. Note that these definitions of argument extensions are not the original ones introduced in [3] but are equivalent formulations [26]. Another way of expressing the semantics of an AA framework is in terms of argument labellings [31,25]. An argument labelling is a total function Lab Arg : Ar → {in, out, undec}. The set of arguments labelled in by Lab Arg is in( Lab Arg ) = { A ∈ Ar | Lab Arg ( A ) = in}; the sets of arguments labelled out and undec are denoted out( Lab Arg ) and undec( Lab Arg ), respectively. An argument labelling Lab Arg is an admissible argument labelling if and only if for each argument A ∈ Ar it holds that:

• if Lab Arg ( A ) = in then for each B ∈ Ar attacking A, Lab Arg ( B ) = out; • if Lab Arg ( A ) = out then there exists some B ∈ Ar attacking A such that Lab Arg ( B ) = in. An argument labelling Lab Arg is a complete argument labelling if and only if it is an admissible argument labelling and for each argument A ∈ Ar it holds that:

• if Lab Arg ( A ) = undec then there exists some B ∈ Ar attacking A such that Lab Arg ( B ) = undec and there exists no C ∈ Ar attacking A such that Lab Arg (C ) = in. In order to define argument labellings according to other semantics, we first recall how to compare the commitment of argument labellings [26]. Let Lab Arg 1 and Lab Arg 2 be argument labellings. Lab Arg 2 is more or equally committed than Lab Arg 1 , denoted Lab Arg 1  Lab Arg 2 , if and only if in( Lab Arg 1 ) ⊆ in( Lab Arg 2 ) and out( Lab Arg 1 ) ⊆ out( Lab Arg 2 ). A complete argument labelling Lab Arg is [25,32]

• a grounded argument labelling if and only if in( Lab Arg ) is minimal (w.r.t. ⊆) among all complete argument labellings; • a preferred argument labelling if and only if in( Lab Arg ) is maximal (w.r.t. ⊆) among all complete argument labellings; • an ideal argument labelling if and only if Lab Arg is a maximal (w.r.t. ) admissible argument labelling which satisfies that for all preferred argument labellings Lab Arg , Lab Arg  Lab Arg ; • a semi-stable argument labelling if and only if undec( Lab Arg ) is minimal (w.r.t. ⊆) among all complete argument labellings;

• a stable argument labelling if and only if undec( Lab Arg ) = ∅. Complete, grounded, preferred, ideal, semi-stable, and stable argument extensions correspond one-to-one to the sets of arguments labelled in by the complete, grounded, preferred, ideal, semi-stable, and stable argument labellings, respectively [25,32]. In contrast, an admissible argument extension may correspond to various admissible argument labellings [25]. 2.2. Assumption-based argumentation An Assumption-Based Argumentation (ABA) framework [4,9,10] is a tuple L, R, A, ¯, where

• (L, R) is a deductive system, with L a language of countably many sentences and R a set of inference rules of the form s0 ← s1 , . . . , sn (n ≥ 0) where s0 , . . . , sn ∈ L; s0 is the head of the inference rule and s1 , . . . , sn is its body; • A ⊆ L is a non-empty set of assumptions; • ¯ is a total mapping from A into L defining the contrary of assumptions, where α denotes the contrary of α ∈ A. An ABA framework is flat if assumptions only occur in the body of inference rules [33]. From here onwards, until Section 6, we assume as given a flat ABA framework L, R, A, ¯. An argument [9] with conclusion s ∈ L and premises Asms ⊆ A, denoted Asms  s, is a finite tree where every node holds a sentence in L or the sentence τ (where τ ∈ / L stands for “true”), such that

• the root node holds s; • for every node N – if N is a leaf then N holds either an assumption or

τ;

C. Schulz, F. Toni / International Journal of Approximate Reasoning 84 (2017) 110–149

115

– if N is not a leaf and N holds the sentence s0 , then there is an inference rule s0 ← s1 , . . . , sm ∈ R and either m = 0 and the only child node of N holds τ or m > 0 and N has m children holding s1 , . . . , sm ; • Asms is the set of all assumptions held by leaf nodes. We sometimes name arguments with capital letters, e.g. A : Asms  s is an argument with name A. With an abuse of notation, the name of an argument is also used to refer to the whole argument. Note that for every assumption α ∈ A there exists an assumption-argument {α }  α . Let Asms, Asms1 , Asms2 ⊆ A be sets of assumptions and let α ∈ A be an assumption.

• Asms attacks α if and only if there exists an argument Asms  α such that Asms ⊆ Asms. Equivalently, we say that α • • • •

is attacked by Asms. Asms1 attacks Asms2 if and only if Asms1 attacks some α ∈ Asms2 . Asms+ = {α ∈ A | Asms attacks α }. Asms defends α if and only if Asms attacks all sets of assumptions attacking Asms is conflict-free if and only if Asms ∩ Asms+ = ∅.

α.

The semantics of an ABA framework are defined as assumption extensions, i.e. sets of accepted assumptions [4,23,24]. A set of assumptions Asms ⊆ A is

• • • • •

an admissible assumption extension if and only if Asms is conflict-free and defends every α ∈ Asms; a complete assumption extension if and only if Asms is conflict-free and consists of all assumptions it defends; a grounded assumption extension if and only if Asms is a minimal (w.r.t. ⊆) complete assumption extension; a preferred assumption extension if and only if Asms is a maximal (w.r.t. ⊆) complete assumption extension; an ideal assumption extension if and only if Asms is a maximal (w.r.t. ⊆) complete assumption extension satisfying that for all preferred assumption extensions Asms , Asms ⊆ Asms ; • a semi-stable assumption extension if and only if Asms is a complete assumption extension and for all complete assumption extensions Asms , Asms ∪ Asms+ ⊂ Asms ∪ Asms + ; • a stable assumption extension if and only if Asms is a complete assumption extension and Asms ∪ Asms+ = A. Note that some of these definitions are not the original ones introduced in [4,23] but are equivalent formulations as proven in [24]. 2.3. Correspondence between ABA and AA A flat ABA framework L, R, A, ¯ can be mapped onto a corresponding AA framework  Ar A B A , Att A B A  [23], where

• Ar A B A is the set of all arguments Asms  s; • ( Asms1  s1 , Asms2  s2 ) ∈ Att A B A if and only if ∃α ∈ Asms2 such that s1 = α . Given an admissible/complete/grounded/preferred/ideal/stable assumption extension Asms of L, R, A, ¯, the set of all arguments whose premises are a subset of Asms is an admissible/complete/grounded/preferred/ideal/stable argument extension of  Ar A B A , Att A B A  [23,20,24]. Conversely, given an admissible/complete/grounded/preferred/ideal/stable argument extension Args of  Ar A B A , Att A B A , the union of all premises of arguments in Args is an admissible/complete/grounded/preferred/ideal/stable assumption extension of L, R, A, ¯ [23,20,24]. Note that this correspondence does not hold for semistable assumption and argument extensions. 3. Assumption labellings Inspired by argument labellings for AA frameworks, we build upon recently introduced labellings for ABA frameworks [27]. In contrast to labellings in AA, which assign labels to whole arguments, assumption labellings assign labels to single assumptions. The three labels used throughout this paper are in, indicating that an assumption is accepted, out, indicating that an assumption is rejected, and undec, indicating that an assumption is neither accepted nor rejected and thus undecided. Definition 1. An assumption labelling is a total function Lab Asm : A → {in, out, undec}. If Lab Asm(α ) = in, we say that α is labelled in by Lab Asm, or equivalently that Lab Asm labels α (as) in. Analogous terminology is used for assumptions labelled out and undec. The set of all assumptions labelled in by Lab Asm is in( Lab Asm) = {α ∈ A | Lab Asm(α ) = in}, and the sets of all assumptions labelled out and undec are denoted out( Lab Asm) and undec( Lab Asm), respectively.

116

C. Schulz, F. Toni / International Journal of Approximate Reasoning 84 (2017) 110–149

We frequently represent an assumption labelling as a set of ordered pairs, for example for an ABA framework with A = {φ, ψ, χ } and the labelling Lab Asm such that Lab Asm(φ) = in, Lab Asm(ψ) = out, and Lab Asm(χ ) = undec, we equivalently represent Lab Asm as {(φ, in), (ψ, out), (χ , undec)}. 3.1. Admissible semantics An admissible assumption extension is a set of accepted assumptions which is able to defend itself. In other words, if an assumption α is contained in an admissible assumption extension, then all sets of assumptions attacking α contain some assumption attacked by this admissible assumption extension. In an admissible assumption labelling the concept of defence is mirrored by requiring that if an assumption α is accepted (labelled in) then all sets of assumptions attacking α contain a rejected assumption (labelled out), which in turn is attacked by a set of accepted assumptions (all labelled in). In addition, we require that an undecided assumption (labelled undec) is not attacked by a set of accepted assumptions (all labelled in), since an assumption attacked by accepted assumptions can clearly not be accepted (due to the conflict-freeness property of the admissible semantics) and should thus be rejected rather than undecided. Definition 2. Let Lab Asm be an assumption labelling. Lab Asm is an admissible assumption labelling if and only if for each assumption α ∈ A it holds that:

• if Lab Asm(α ) = in then for each set of assumptions Asms attacking α there exists some β ∈ Asms such that Lab Asm(β) = out; • if Lab Asm(α ) = out then there exists a set of assumptions Asms attacking α such that for all β ∈ Asms, Lab Asm(β) = in;

• if Lab Asm(α ) = undec then for each set of assumptions Asms attacking α there exists some β ∈ Asms such that Lab Asm(β) = in. Example 1. Consider the following ABA framework, which we call A B A 1 :

L = {r , p , x, ρ , ψ, χ }, R = { p ← ρ ; x ← ψ}, A = {ρ , ψ, χ }, ρ = r , ψ = p , χ = x. A B A 1 has three admissible assumption labellings:

• Lab Asm1 = {(ρ , undec), (ψ, undec), (χ , undec)}, • Lab Asm2 = {(ρ , in), (ψ, out), (χ , undec)}, and • Lab Asm3 = {(ρ , in), (ψ, out), (χ , in)}. These assumption labellings demonstrate two important points: first, an assumption which is not attacked by any set of assumption (ρ in A B A 1 ) cannot be labelled out; second, an assumption attacked by a set of assumptions containing only in labelled assumptions (ψ in Lab Asm2 and Lab Asm3 ) is always labelled out. It is important to note that the empty set of assumptions has a special role as an attacking set of assumptions: any assumption attacked by the empty set is labelled out by all admissible assumption labellings since the empty set stands for a (non-refutable) fact, so the attacked assumption clearly has to be rejected, as illustrated in Example 2. Example 2. Let A B A 2 be A B A 1 from Example 1 with the additional sentences φ and f in L, where φ is an assumption with φ = f , and with the additional inference rule f ←. Since {}  f is an argument, φ is attacked by the empty set of assumptions as well as by all other sets of assumptions. Thus, φ cannot be labelled in since the attacking empty set does not contain an assumption labelled out, and φ cannot be labelled undec since the attacking empty set does not contain an assumption not labelled in. Consequently, φ is labelled out by all admissible assumption labellings. A B A 2 has thus three admissible assumption labellings:

• Lab Asm1 = {(φ, out), (ρ , undec), (ψ, undec), (χ , undec)} • Lab Asm2 = {(φ, out), (ρ , in), (ψ, out), (χ , undec)} • Lab Asm3 = {(φ, out), (ρ , in), (ψ, out), (χ , in)} Note that these are the same admissible assumption labellings as for A B A 1 , but with the additional assumption φ which is always labelled out. The number of admissible assumption labellings is thus not influenced by assumptions attacked by the empty set since these assumptions do not have alternative labels in different admissible assumption labellings.

C. Schulz, F. Toni / International Journal of Approximate Reasoning 84 (2017) 110–149

117

The following theorem shows that there is a one-to-one correspondence between the admissible semantics in terms of assumption labellings and extensions. Theorem 1. 1. Let Asms be an admissible assumption extension. Then Lab Asm with in( Lab Asm) = Asms, out( Lab Asm) = Asms+ , and undec( Lab Asm) = A \ ( Asms ∪ Asms+ ) is an admissible assumption labelling. 2. Let Lab Asm be an admissible assumption labelling. Then Asms = in( Lab Asm) is an admissible assumption extension with Asms+ = out( Lab Asm) and A \ ( Asms ∪ Asms+ ) = undec( Lab Asm). The proof of Theorem 1 as well as all other results can be found in the Appendix. Example 3. A B A 2 from Example 2 has three admissible assumption extensions: Asms1 = {}, Asms2 = {ρ }, and Asms3 = {ρ , χ }, corresponding to the three admissible assumption labellings Lab Asm1 , Lab Asm2 , and Lab Asm3 , respectively. Note that without the third condition in Definition 2, the second item in Theorem 1 would not hold. For example, Lab Asm4 = {(φ, out), (ρ , in), (ψ, undec), (χ , undec)} would be an admissible assumption labelling of A B A 2 (see Example 2), but even though Asms4 = in( Lab Asm4 ) = {ρ } is an admissible assumption extension of A B A 2 , it does not hold that + Asms+ / out( Lab Asm4 ). 4 = out( Lab Asm4 ) as stated in the second item of Theorem 1 since ψ ∈ Asms4 but ψ ∈ If an assumption is defended by an admissible assumption extension then adding this assumption to the extension yields another admissible assumption extension [4] (similar to the Fundamental Lemma in AA [3]). Due to the one-to-one correspondence between admissible assumption labellings and extensions, an analogue property holds for admissible assumption labellings. Lemma 2. Let Lab Asm be an admissible assumption labelling and let α be such that for each set of assumptions Asms attacking α there exists some β ∈ Asms such that Lab Asm(β) = out. Let α = {γ ∈ A | ∃ Asms ⊆ A such that α ∈ Asms, Asms attacks γ , ∀δ ∈ Asms : δ = α → Lab Asm(δ) = in}. Then Lab Asm with in( Lab Asm ) = in( Lab Asm) ∪ {α }, out( Lab Asm ) = out( Lab Asm) ∪ α and undec( Lab Asm ) = undec( Lab Asm) \ ({α } ∪ α ) is an admissible assumption labelling. The lemma states that if an assumption α is defended by an admissible assumption labelling, i.e. all attacking sets of assumptions Asms contain an assumption β labelled out, then changing the label of α to in and changing the label of all assumptions γ which now need to be rejected (due to the change of label of α ) to out yields another admissible assumption labelling. Example 4. Let A B A 3 be an ABA framework with:

L = { f , p , r , x, φ, ψ, ρ , χ }, R = {r ← φ, χ ; p ← ρ }, A = {φ, ψ, ρ , χ }, φ = f , ψ = p, ρ = r, χ = x. Lab Asm1 = {(φ, undec), (ψ, undec), (ρ , undec), (χ , in)} is an admissible assumption labellings of A B A 3 . Since φ is not attacked by any set of assumptions, it holds that each set of assumptions attacking φ contains an assumption labelled out, and φ + = {ρ }. As stated in Lemma 2, Lab Asm2 with in( Lab Asm2 ) = {χ , φ}, out( Lab Asm2 ) = {ρ }, and undec( Lab Asm2 ) = {ψ} is an admissible assumption labelling of A B A 3 . Since with respect to Lab Asm2 it holds that each set of assumptions attacking ψ contains an assumption labelled out, Lab Asm3 with in( Lab Asm3 ) = {χ , φ, ψ}, out( Lab Asm3 ) = {ρ }, and undec( Lab Asm3 ) = {} is also an admissible assumption labelling of A B A 3 . 3.2. Complete semantics In addition to defending itself against attackers, a complete assumption extension contains every assumption it defends. This additional condition is mirrored in complete assumption labellings by requiring that an assumption which is defended has to be labelled in. This can be achieved by modifying the definition of admissible assumption labellings in various ways. In an admissible assumption labelling a defended assumption may be labelled in or undec. Thus, one way of modifying the definition of admissible assumption labellings is to prohibit labelling defended assumptions as undec. In other words, an assumption labelled undec has to be attacked by at least one set of assumptions which does not contain any assumption labelled out. Definition 3. Let Lab Asm be an assumption labelling. Lab Asm is a complete assumption labelling if and only if Lab Asm is an admissible assumption labelling and for each assumption α ∈ A it holds that:

118

C. Schulz, F. Toni / International Journal of Approximate Reasoning 84 (2017) 110–149

• if Lab Asm(α ) = undec then there exists a set of assumptions Asms attacking α such that for all γ ∈ Asms, Lab Asm(γ ) = out. Note that the new condition for undec assumptions implies that there exists a set of assumptions Asms attacking α such that for some γ ∈ Asms, Lab Asm(γ ) = undec, since by the definition of admissible assumption labellings some β ∈ Asms is not labelled in and by the new condition no γ ∈ Asms is labelled out. However, the fact that there exists a set of assumptions Asms attacking α such that for some γ ∈ Asms, Lab Asm(γ ) = undec, is not a sufficient condition for characterising admissible assumption labellings which are also complete assumption labellings. Example 5. Consider again A B A 1 from Example 1 and its three admissible assumption labellings. In Lab Asm1 , ρ does not satisfy the new condition for undec assumptions, and in Lab Asm2 , χ does not satisfy the new condition. The only admissible assumption labelling satisfying the new condition is Lab Asm3 , which is thus the only complete assumption labelling of A B A 1 . The second way to modify the definition of admissible assumption labellings to express the complete semantics is to add a condition which explicitly states that if an assumption α is defended, i.e. if all sets of assumptions attacking α contain some assumption labelled out, then α has to be labelled in. This condition adds the “opposite direction” of the first condition of an admissible assumption labelling. To make this way of defining complete assumption labellings more uniform, the “opposite direction” of the second condition of an admissible assumption labelling is added, too. This renders the third condition of an admissible assumption labelling superfluous and thus leaves two “if and only if” conditions to be satisfied by each α ∈ A:

• Lab Asm(α ) = in if and only if for each set of assumptions Asms attacking α there exists some β ∈ Asms such that Lab Asm(β) = out; • Lab Asm(α ) = out if and only if there exists a set of assumptions Asms attacking α such that for all β ∈ Asms, Lab Asm(β) = in. Since Lab Asm is an assumption labelling and thus labels each assumption in an ABA framework, assumptions which do not satisfy the right hand side of either of the above conditions are “automatically” labelled undec by Lab Asm. A third way to define complete assumption labellings reverses all three conditions of Definition 3, thus specifying which label an assumption satisfying a certain condition should have. Theorem 3. Let Lab Asm be an assumption labelling. The following statements are equivalent: 1. Lab Asm is a complete assumption labelling. 2. Lab Asm is such that for each α ∈ A it holds that: • Lab Asm(α ) = in if and only if for each set of assumptions Asms attacking α there exists some β ∈ Asms such that Lab Asm(β) = out; • Lab Asm(α ) = out if and only if there exists a set of assumptions Asms attacking α such that for all β ∈ Asms, Lab Asm(β) = in. 3. Lab Asm is such that for each α ∈ A it holds that: • if for each set of assumptions Asms attacking α there exists some β ∈ Asms such that Lab Asm(β) = out, then Lab Asm(α ) = in; • if there exists a set of assumptions Asms attacking α such that for all β ∈ Asms, Lab Asm(β) = in, then Lab Asm(α ) = out; • if for each set of assumptions Asms1 attacking α there exists some β ∈ Asms1 such that Lab Asm(β) = in, and there exists a set of assumptions Asms2 attacking α such that for all γ ∈ Asms2 , Lab Asm(γ ) = out, then Lab Asm(α ) = undec. Example 6. Consider again A B A 1 from Example 1 and its three admissible assumption labellings. Lab Asm1 does not satisfy the second item in Theorem 3 since ρ violates the first condition. Similarly, Lab Asm1 does not satisfy the third item in Theorem 3 since ρ violates the first condition. Lab Asm2 does not satisfy the second or third item in Theorem 3 since χ violates the first condition of both items. Only Lab Asm3 satisfies the second as well as the third item in Theorem 3, and is thus the only complete assumption labelling of A B A 1 . All three ways of defining complete assumption labellings are useful in their own rights. Definition 3 is particularly suitable to verify whether a given assumption labelling is indeed a complete assumption labelling. In contrast, the third item in Theorem 3 is more suitable for determining which assumptions should have which label. Since the second item in Theorem 3 can be considered as the “union” of the two other definitions, it lends itself to either of the two tasks. Note that the definition of admissible assumption labellings cannot be equivalently expressed by reversing of the conditions in Definition 2 since they are not mutually exclusive. In particular, an unattacked assumption would satisfy both the condition to be labelled in and to be labelled undec, so not matter which of the two labels was assigned to the assumption, one of the two conditions would be violated.

C. Schulz, F. Toni / International Journal of Approximate Reasoning 84 (2017) 110–149

119

The following theorem proves that there is a one-to-one correspondence between complete assumption labellings and extensions, just as between admissible assumption labellings and extensions. Theorem 4. 1. Let Asms be a complete assumption extension. Then Lab Asm with in( Lab Asm) = Asms, out( Lab Asm) = Asms+ , and undec( Lab Asm) = A \ ( Asms ∪ Asms+ ) is a complete assumption labelling. 2. Let Lab Asm be a complete assumption labelling. Then Asms = in( Lab Asm) is a complete assumption extension with Asms+ = out( Lab Asm) and A \ ( Asms ∪ Asms+ ) = undec( Lab Asm). 3.3. Grounded, preferred, ideal, semi-stable, and stable semantics Based on the notion of complete assumption labellings, the grounded, preferred, ideal, semi-stable, and stable semantics can be defined in terms of assumption labellings. Definition 4. A complete assumption labelling Lab Asm is:

• a grounded assumption labelling if and only if in( Lab Asm) is minimal (w.r.t. ⊆) among all complete assumption labellings; • a preferred assumption labelling if and only if in( Lab Asm) is maximal (w.r.t. ⊆) among all complete assumption labellings; • an ideal assumption labelling if and only if in( Lab Asm) is maximal (w.r.t. ⊆) among all complete assumption labellings satisfying that for all preferred assumption labellings Lab Asm , in( Lab Asm) ⊆ in( Lab Asm ); • a semi-stable assumption labelling if and only if undec( Lab Asm) is minimal (w.r.t. ⊆) among all complete assumption labellings;

• a stable assumption labelling if and only if undec( Lab Asm) = ∅. Example 7. Let A B A 4 be the ABA framework with:

L = {r , p , x, ρ , ψ, χ }, R = {r ← ψ ; p ← ρ ; p ← χ ; x ← ψ ; x ← χ }, A = {ρ , ψ, χ }, ρ = r , ψ = p , χ = x. A B A 4 has three complete assumption labellings:

• Lab Asm1 = {(ρ , undec), (ψ, undec), (χ , undec)}, • Lab Asm2 = {(ρ , out), (ψ, in), (χ , out)}, and • Lab Asm3 = {(ρ , in ), (ψ, out), (χ , undec)}. Lab Asm1 is the grounded assumption labelling, Lab Asm2 and Lab Asm3 are both preferred assumption labellings, Lab Asm1 is the ideal assumption labelling, and Lab Asm2 is the only stable as well as the only semi-stable assumption labelling. The following theorem proves that the grounded, preferred, ideal, semi-stable, and stable assumption labellings correspond one-to-one to the respective assumption extensions. Theorem 5. 1. Let Asms be a grounded/preferred/ideal/semi-stable/stable assumption extension. Then Lab Asm with in( Lab Asm) = Asms, out( Lab Asm) = Asms+ , and undec( Lab Asm) = A \ ( Asms ∪ Asms+ ) is a grounded/preferred/ideal/semi-stable/stable assumption labelling. 2. Let Lab Asm be a grounded/preferred/ideal/semi-stable/stable assumption labelling. Then Asms = in( Lab Asm) is a grounded/preferred/ideal/semi-stable/stable assumption extension with Asms+ = out( Lab Asm) and A \ ( Asms ∪ Asms+ ) = undec( Lab Asm). Corollary 6 follows straightaway from the correspondence of grounded and ideal assumption labellings and extensions and the uniqueness of grounded and ideal assumption extensions [4,23]. Corollary 6. The grounded and ideal assumption labelling are both unique. We now show that preferred, ideal, semi-stable, and stable assumption labellings can be redefined in terms of admissible (rather than complete) assumption labellings.

120

C. Schulz, F. Toni / International Journal of Approximate Reasoning 84 (2017) 110–149

Proposition 7. Let Lab Asm be an admissible assumption labelling.

• Lab Asm is a preferred assumption labelling if and only if in( Lab Asm) is maximal (w.r.t. ⊆) among all admissible assumption labellings.

• Lab Asm is an ideal assumption labelling if and only if in( Lab Asm) is maximal (w.r.t. ⊆) among all admissible assumption labellings satisfying that for all preferred assumption labellings Lab Asm , in( Lab Asm) ⊆ in( Lab Asm ). • Lab Asm is a semi-stable assumption labelling if and only if undec( Lab Asm) is minimal (w.r.t. ⊆) among all admissible assumption labellings.

• Lab Asm is a stable assumption labelling if and only if undec( Lab Asm) = ∅. Example 8. The results from Proposition 7 are illustrated by A B A 1 (see Examples 1 and 5). For example, the only maximal admissible assumption labelling of A B A 1 is Lab Asm3 , which is also the only maximal complete, and thus preferred, assumption labelling of A B A 1 . 4. Argument-supporting sets of assumptions Determining admissible or complete assumption labellings of an ABA framework as well as checking whether an assumption labelling is admissible or complete can be cumbersome since some conditions specifying the label of an assumption α require to consider every set of assumptions attacking α . In particular, not only the set of premises of an argument whose conclusion is the contrary of α attacks α , but also every superset thereof. Example 9. To verify whether χ is correctly labelled in the admissible assumption labelling Lab Asm3 = {(ρ , in), (ψ, out), (χ , in)} of A B A 1 (see Example 1), not only the set of assumptions {ψ}, which forms the premises of an argument with conclusion x (the contrary of χ ), but also every superset thereof, i.e. {ρ , ψ}, {ψ, χ }, and {ρ , ψ, χ }, has to be checked. In this section, we show that considering only sets of assumptions which form the premises of some argument, which we call argument-supporting sets of assumptions, when determining or checking assumption labellings is equivalent to considering all sets of assumptions. This is inspired by the fact that assumption extensions can be determined and checked by considering either all or only argument-supporting sets of assumptions [33]. 4.1. Assumption labellings with respect to argument-supporting sets of assumptions A set of assumptions is argument-supporting if it forms the premises of some argument. Definition 5. Let Asms ⊆ A be a set of assumptions. Asms is an argument-supporting set of assumptions if and only if there exists some s ∈ L such that Asms  s is an argument. Note that all singleton sets of assumptions are argument-supporting, i.e. for every assumption supporting set of assumptions since {α }  α is an argument.

α ∈ A, {α } is an argument-

Notation 6. The set of all argument-supporting sets of assumptions is Sarg = { Asms ⊆ A | Asms is an argument-supporting set of assumptions}. We define a variant of admissible assumption labellings where only argument-supporting, rather than all, sets of assumptions attacking an assumption are taken into account. Definition 7. Let Lab Asm be an assumption labelling. Lab Asm is an admissible assumption labelling with respect to argumentsupporting sets if and only if for each assumption α ∈ A it holds that:

• if Lab Asm(α ) = in then for each argument-supporting set of assumptions Asms attacking α there exists some β ∈ Asms such that Lab Asm(β) = out; • if Lab Asm(α ) = out then there exists an argument-supporting set of assumptions Asms attacking α such that for all β ∈ Asms, Lab Asm(β) = in; • if Lab Asm(α ) = undec then for each argument-supporting set of assumptions Asms attacking α there exists some β ∈ Asms such that Lab Asm(β) = in. To check whether χ in A B A 1 is correctly labelled according to admissible assumption labellings with respect to argument-supporting sets, only the set {ψ} has to be taken into account (compare Example 9). The following Lemma shows that the original definition of admissible assumption labellings (Definition 2) and the new definition of admissible assumption labellings with respect to argument-supporting sets can be used interchangeably. This

C. Schulz, F. Toni / International Journal of Approximate Reasoning 84 (2017) 110–149

121

extends the result of [33] that admissible assumption extensions can be equivalently defined in terms of all sets of assumptions or argument-supporting sets of assumptions. Lemma 8. Let Lab Asm be an assumption labelling. Lab Asm is an admissible assumption labelling if and only if Lab Asm is an admissible assumption labelling with respect to argument-supporting sets. Analogously to admissible assumption labellings, we define a variant of complete assumption labellings, where only argument-supporting sets of assumptions attacking an assumption in question are taken into account. Definition 8. Let Lab Asm be an assumption labelling. Lab Asm is a complete assumption labelling with respect to argumentsupporting sets if and only if Lab Asm is an admissible assumption labelling with respect to argument-supporting sets and for each assumption α ∈ A it holds that:

• if Lab Asm(α ) = undec then there exists an argument-supporting set of assumptions Asms attacking α such that for all γ ∈ Asms, Lab Asm(γ ) = out. As for admissible assumption labellings, the notions of complete assumption labellings and complete assumption labellings with respect to argument-supporting sets are equivalent. Proposition 9. Let Lab Asm be an assumption labelling. Lab Asm is a complete assumption labelling if and only if Lab Asm is a complete assumption labelling with respect to argument-supporting sets. Since the grounded, preferred, ideal, semi-stable, and stable assumption labellings are based on complete/admissible assumption labellings, it follows that they can be equivalently defined in terms of complete/admissible assumption labellings with respect to argument-supporting sets. Note that here we do not address the issue of computing argument-supporting sets of assumptions, as this is beyond the scope of this paper. However, we note that in order to determine complete assumption labellings, arguments have to be constructed when considering argument-supporting sets as well as when considering all sets of assumptions. Once arguments have been constructed for all sets of assumptions or argument-supporting sets, using argument-supporting sets may be beneficial for determining complete assumption labellings. The reason is that, depending on the ABA framework and in particular on the set of inference rules R, the set of all argument-supporting sets of assumptions Sarg may be much smaller than or equal to the set of all sets of assumptions ℘ (A). For example, in A B A 1 from Example 1 the set of all argument-supporting sets of assumptions consists only of the singleton sets, i.e. {{ρ }, {ψ}, {χ }}, whereas the set of all sets of assumptions is ℘ ({ρ , ψ, χ }) = {{}, {ρ }, {ψ}, {χ }, {ρ , ψ}, {ρ , χ }, {ψ, χ }, {ρ , ψ, χ }}. Therefore, considering only argument-supporting sets of assumptions when determining complete assumption labellings may in the best case require to check only a fraction of all sets of assumptions, but in the worst case it is exactly the same. Observation 10. Let Sall = ℘ (A) be the set of all sets of assumptions, so |Sall | = 2|A| .

• In the best case, |Sarg | = |A|. This is for example the case if R = ∅, since the only argument-supporting sets of assumptions are the singleton sets.

• In the worst case, |Sarg | = |Sall | = 2|A| . This is for example the case if R is such that for each Asmsi ∈ Sall there exists some inference rule si ← si 1 , . . . , sin ∈ R such that Asmsi = {si 1 , . . . , sin }. Example 10. Let A B A 5 be the ABA framework with:

L = { p , r , ψ, ρ }, R = { p ← ; p ← ρ ; r ← ψ; r ← ψ, ρ }, A = {ψ, ρ }, ψ = p , ρ = r. Here, the set of all argument-supporting sets of assumptions is Sarg = {{}, {ψ}, {ρ }, {ψ, ρ }}, which is the same as the set of all sets of assumptions. 4.2. ABA graphs In most of the ABA literature (e.g. [33,23,12,34,10,35]), ABA frameworks are not displayed graphically; they are simply given as tuples, as done in the Examples presented so far. We introduce ABA graphs, where nodes are argument-supporting sets of assumptions and edges are attacks between these argument-supporting sets of assumptions.

122

C. Schulz, F. Toni / International Journal of Approximate Reasoning 84 (2017) 110–149

Fig. 5. Left: the ABA graph of A B A 1 . Right: the graph illustrating all sets of assumptions of A B A 1 and all attacks between them.

Fig. 6. The ABA graph of A B A 4 (see Example 11). The right version also indicates one of the complete assumption labellings of A B A 4 .

Definition 9. The ABA graph G = ( V , E ) is a directed graph with V = Sarg and E = {( Asms1 , Asms2 ) | Asms1 , Asms2 ∈ V and Asms1 attacks Asms2 }. The ABA graph of A B A 1 from Example 1 has only three nodes, namely the singleton sets of assumptions, as shown on the left of Fig. 5. As a comparison, the right of Fig. 5 illustrates the graph of all sets of assumptions and attacks between them. Since an ABA graph illustrates all argument-supporting sets of assumptions and attacks between them, an ABA graph can be used to determine the semantics of an ABA framework. Example 11. The ABA graph of A B A 4 (see Example 7) is displayed on the left of Fig. 6. It illustrates which (argumentsupporting) sets of assumptions have to be taken into account when determining complete or admissible assumption labellings (with respect to argument-supporting sets). For example, for ρ to be labelled in by a complete assumption labelling (with respect to argument-supporting sets), all (argument-supporting) sets of assumptions attacking ρ have to contain an assumption labelled out. Since the only set of assumptions attacking ρ in the ABA graph is {ψ}, we deduce that ψ has to be labelled out by any complete assumption labelling that labels ρ as in. It is then easy to verify, based on the two sets of assumptions attacking χ , that with ψ labelled out, χ can only be labelled undec. This complete assumption labelling of A B A 4 is illustrated in the ABA graph on the right of Fig. 6. Another graphical representation of an ABA framework is the attack relationship graph [4], which was introduced to characterise different types of ABA frameworks. The question thus arises whether attack relationship graphs can also be used to determine the semantics, in particular the assumption labellings, of an ABA framework. The attack relationship graph Gatt = ( V , E ) is a directed graph with V = A and E = {(α , β) | such that Asms attacks β and  Asms ⊂ Asms such that Asms attacks β}.

α , β ∈ V and α ∈ Asms

The main difference between an ABA graph and an attack relationship graph is that the vertices of an ABA graph are sets of assumptions, including all the singleton sets, whereas the vertices of an attack relationship graph are single assumptions. The following example demonstrates that attack relationship graphs do not capture enough information to determine the semantics of an ABA framework. Example 12. Let A B A 6 be the ABA framework with:

L = {ρ , ψ, χ , φ, ω, r , p , x}, R = {r ← φ ; r ← ω ; r ← ψ, χ }, A = {ρ , ψ, χ , φ, ω}, ρ = r, ψ = p, χ = x, φ = ψ , ω = ψ . Furthermore, let A B A 7 have the same L, A, and contraries as A B A 6 , but with R = {r ← φ ; r ← ψ, ω ; r ← χ , ω}. The ABA graphs of A B A 6 and A B A 7 , which are structurally different, are displayed in Fig. 7. In contrast, the attack relationship graphs of A B A 6 and A B A 7 are the same, as illustrated in Fig. 8. Thus, based on the attack relationship graph it is impossible to distinguish A B A 6 and A B A 7 . However, the two ABA frameworks have different complete labellings, as indicated in Fig. 7.

C. Schulz, F. Toni / International Journal of Approximate Reasoning 84 (2017) 110–149

123

Fig. 7. The ABA graphs of A B A 6 (left) and A B A 7 (right) from Example 12, each with their only complete assumption labelling.

Fig. 8. The attack relationship graph of both A B A 6 and A B A 7 from Example 12.

Fig. 9. The attack relationship graph (left) and the ABA graph (right) of A B A 8 from Example 13.

It is therefore not possible to determine the complete or admissible assumption labellings of A B A 6 and A B A 7 , and more generally of any ABA framework, based on their attack relationship graphs. Conversely, ABA graphs cannot be (straightforwardly) used for the purpose of attack relationship graphs, i.e. to characterise different types of ABA frameworks. For example, an ABA framework is stratified if and only if its attack relationship graph does not have an infinite sequence of edges [4]. However, ABA graphs may have an infinite sequence of edges even if the attack relationship graph does not, as demonstrated in Example 13. Example 13. Let A B A 8 be the following ABA framework:

L = { p , r , x, ψ, ρ , χ }, R = {r ← ψ ; x ← ρ ; r ← ψ, χ }, A = {ψ, ρ , χ }, ψ = p , ρ = r , χ = x. The attack relationship graph and the ABA graph of A B A 8 are displayed in Fig. 9. Since the attack relationship graph does not have any infinite sequence of edges, A B A 8 is stratified. However, the ABA graph does have an infinite sequence of edges since it comprises a cycle. The example illustrates that an infinite sequence of edges in an ABA graph does not indicate that the ABA framework is not stratified. Proposition 11. Let L, R, A, ¯ be an ABA framework and let G be its ABA graph and Gatt its attack relationship graph. If Gatt has an infinite sequence of edges then G has an infinite sequence of edges, but not vice versa. Another way to graphically represent an ABA framework is in terms of the AA graph of its corresponding AA framework, as for example done in [24,22]. Interestingly, even though nodes in an ABA graph are argument-supporting sets of assumptions, an ABA graph does not generally have the same number of nodes as the AA graph of the corresponding AA framework, where nodes are arguments. In particular, an AA graph may have more nodes than an ABA graph since the same set of assumptions may form the set of premises of various arguments. As an example, compare the ABA graph of A B A 4 shown in Fig. 6 with the AA graph of its corresponding AA framework illustrated in Fig. 10.

124

C. Schulz, F. Toni / International Journal of Approximate Reasoning 84 (2017) 110–149

Fig. 10. AA graph of the corresponding AA framework of A B A 4 .

Recently a new way to represent arguments of an ABA framework has been introduced with the purpose of eliminating redundancies in arguments [36], namely as argument graphs rather than tree-structured arguments. Various argument graphs can furthermore be combined to form a larger argument graph, which represents a set of arguments without redundancies. Since the semantics of ABA frameworks in terms of argument graphs is slightly different from the semantics of ABA frameworks in terms of assumption and argument extensions [36], a detailed comparison between argument graphs and ABA graphs goes beyond the scope of this paper. 5. Assumption labellings versus argument labellings In this section, we examine the relationship between assumption labellings of an ABA framework and argument labellings of its corresponding AA framework. In the remainder, and if clear from the context, we assume as given a flat ABA framework L, R, A, ¯ and its corresponding AA framework  Ar A B A , Att A B A  (see Section 2). 5.1. Translating between assumption and argument labellings Before going into detail about the (non-)correspondence between assumption and argument labellings according to the various semantics, we examine the relationship between assumption and argument labellings in general. Notation 10. L Asm denotes the set of all assumption labellings of L, R, A, ¯ and L Arg the set of all argument labellings of  Ar A B A , Att A B A . First, we observe that the number of all possible assumption labellings of an ABA framework is smaller than or equal to the number of all possible argument labellings of its corresponding AA framework since an assumption labelling labels only assumptions, i.e. |A| elements, whereas an argument labelling labels every assumption-argument as well as every argument constructed using inference rules in R. Observation 12. Since assumption labellings assign one of three labels to each assumption, |L Asm | = 3|A| . Since argument labellings assign one of three labels to each argument, |L Arg | = 3| Ar A B A | .

• In the best case |L Arg | = |L Asm | = 3|A| . This is the case if the only arguments are assumption-arguments, so | Ar A B A | = |A|, for example if R = ∅. • In all other cases |L Arg | > |L Asm |. This is the case if there exists at least one argument which is not an assumption-argument, so | Ar A B A | > |A|, for example if there exists an inference rule s ← ∈ R. As an example, A B A 4 from Example 7 has three assumptions, so there are |L Asm | = 33 = 27 possible assumption labellings. In contrast, the corresponding AA framework of A B A 4 has eight arguments (see Fig. 10), so there are |L Arg | = 38 = 6561 possible argument labellings. We will see in the following sections that even though the number of possible assumption labellings of an ABA framework may be less than the number of possible argument labellings of its corresponding AA framework, the number of assumption and argument labellings according to various semantics is the same. In order to compare assumption and argument labellings, we define two functions for translating between the two types of labellings. The first translation, LabAsm2LabArg, determines the labels of arguments based on the given labels of premises of these arguments. Definition 11. LabAsm2LabArg : L Asm → L Arg maps an assumption labelling Lab Asm to an argument labelling Lab Arg such that:

C. Schulz, F. Toni / International Journal of Approximate Reasoning 84 (2017) 110–149

125

• in( Lab Arg ) = { Asms  s ∈ Ar A B A | Asms ⊆ in( Lab Asm)}, • out( Lab Arg ) = { Asms  s ∈ Ar A B A | ∃α ∈ Asms : α ∈ out( Lab Asm)}, • undec( Lab Arg ) = { Asms  s ∈ Ar A B A | ∃α ∈ Asms : α ∈ undec( Lab Asm), Asms ∩ out( Lab Asm) = ∅}. LabAsm2LabArg mirrors the correspondence between assumption and argument extensions (see Section 2.3) through the mapping from in labelled assumption to in labelled arguments. In addition, LabAsm2LabArg maps out and undec labelled assumptions to out and undec labelled arguments, respectively. An argument is labelled out if one of its premises is labelled out, independently of the labels of its other premises. The intuition of this translation is that an assumption α which is labelled out is attacked by a set of assumptions labelled in. Since this set gives rise to an in labelled argument, any argument which has α as its premise is attacked by an in labelled argument and should thus be labelled out. Arguments labelled undec are simply those whose premises fulfil neither the condition for in nor for out labelled arguments. Lemma 13. LabAsm2LabArg is an injective function but not generally a surjective function. The second translation, LabArg2LabAsm, determines the labels of assumptions based on the given labels of assumption-arguments. Definition 12. LabArg2LabAsm : L Arg → L Asm maps an argument labelling Lab Arg to an assumption labelling Lab Asm such that:

• in( Lab Asm) = {α ∈ A | {α }  α ∈ in( Lab Arg )}, • out( Lab Asm) = {α ∈ A | {α }  α ∈ out( Lab Arg )}, • undec( Lab Asm) = {α ∈ A | {α }  α ∈ undec( Lab Arg )}. In contrast to LabAsm2LabArg, the translation from in labelled arguments to in labelled assumptions in terms of LabArg2LabAsm does not mirror the correspondence between argument and assumption extensions (see Section 2). In particular, the set of in labelled assumptions consists of all assumptions whose assumption-arguments are labelled in, rather than of all assumptions occurring as the premise of any argument labelled in (which would mirror the correspondence between argument and assumption extensions). This is to ensure that the translation of any argument labelling results in a well-defined assumption labelling. Example 14. Let A B A 9 be the ABA framework with:

L = {ρ , ψ, r , p }, R = {r ← ρ }, A = {ρ , ψ}, ρ = ψ , ψ = p. The corresponding AA framework of A B A 9 has three arguments, A 1 : {ρ }  ρ , A 2 : {ψ}  ψ , and A 3 : {ρ }  r. Let Lab Arg be the argument labelling {( A 1 , out), ( A 2 , out), ( A 3 , in)}. Then LabArg2LabAsm( Lab Arg ) = {(ψ, in), (ρ , out)} is a welldefined assumption labelling. However, if the set of assumptions labelled in was defined in such a way that it mirrors the correspondence between argument and assumption extensions, i.e. in( Lab Asm) = {α ∈ A | ∃ Asms  s ∈ in( Lab Arg ), α ∈ Asms}, then ρ ∈ in because A 3 ∈ in( Lab Arg ) but also ρ ∈ out because A 1 ∈ out( Lab Arg ). Note that if LabArg2LabAsm was a function between admissible or complete, rather than arbitrary, argument and assumption labellings, the translation from in labelled arguments to in labelled assumptions could mirror the correspondence between argument and assumption extensions [27]. Lemma 14. LabArg2LabAsm is a surjective function but not generally an injective function. 5.2. Complete semantics Due to the one-to-one correspondence between complete assumption labellings and extensions (Theorem 4), between complete assumption and argument extensions [20,24], and between complete argument extensions and labellings [25], there is also a one-to-one correspondence between complete assumption labellings and complete argument labellings. Theorem 15 below characterises the complete argument labelling corresponding to a given complete assumption labelling in terms of the mapping LabAsm2LabArg. Theorem 15. Let Lab Asm be an assumption labelling. Lab Asm is a complete assumption labelling if and only if

LabAsm2LabArg( Lab Asm) is a complete argument labelling.

126

C. Schulz, F. Toni / International Journal of Approximate Reasoning 84 (2017) 110–149

Note that in addition to proving that every complete assumption labelling is translated to a corresponding complete argument labelling by LabAsm2LabArg, analogous to the correspondence between complete assumption and argument extensions [24], Theorem 15 also proves that for every complete argument labelling which is the translation of some assumption labelling Lab Asm in terms of LabAsm2LabArg, Lab Asm is a complete assumption labelling. Since LabAsm2LabArg is injective but not generally surjective (see Lemma 13), there may be an argument labelling Lab Arg which is not the translation of any assumption labelling in terms of LabAsm2LabArg, so a natural question is whether Lab Arg may be a complete argument labelling. The following Proposition shows that this is not the case, i.e. every complete argument labelling is the translation of some assumption labelling in terms of LabAsm2LabArg . Proposition 16. Let Lab Arg be a complete argument labelling. Then there exists a unique assumption labelling Lab Asm such that LabAsm2LabArg( Lab Asm) = Lab Arg. It follows directly from Theorem 15 that Lab Asm is a complete assumption labelling. We now examine the translation from argument to assumption labellings in terms of LabArg2LabAsm. Theorem 17 below shows that the translation of a complete argument labelling yields a complete assumption labelling. Theorem 17. Let Lab Arg be an argument labelling. If Lab Arg is a complete argument labelling then LabArg2LabAsm( Lab Arg ) is a complete assumption labelling. Note that since LabArg2LabAsm is surjective (see Lemma 14) the converse of Theorem 17 does not hold, i.e. a complete assumption labelling Lab Asm may be the translation of some argument labelling in terms of LabArg2LabAsm which is not a complete argument labelling, as illustrated by the following example. Example 15. A B A 9 from Example 14 has only one complete assumption labelling Lab Asm1 = {(ρ , out), (ψ, in)}. The corresponding AA framework of A B A 9 has three arguments, A 1 : {ρ }  ρ , A 2 : {ψ}  ψ , and A 3 : {ρ }  r, and one complete argument labelling, Lab Arg 1 = {( A 1 , out), ( A 2 , in), ( A 3 , out)}. It holds that LabArg2LabAsm( Lab Arg 1 ) = Lab Asm1 , but also that for Lab Arg 2 = {( A 1 , out), ( A 2 , in), ( A 3 , undec)} LabArg2LabAsm( Lab Arg 2 ) = Lab Asm1 , where Lab Arg 2 is not a complete argument labelling. However, a weaker version of the converse of Theorem 17 holds: Every complete assumption labelling is the translation of some complete argument labelling in terms of LabArg2LabAsm. Lemma 18. Let Lab Asm be a complete assumption labelling. Then there exists a complete argument labelling Lab Arg such that

LabArg2LabAsm( Lab Arg ) = Lab Asm. Even though there may be multiple argument labellings which are translated to the same complete assumption labelling in terms of LabArg2LabAsm, there are no two complete argument labellings which are translated to the same assumption labelling. Lemma 19. Let

Lab Arg 1 = Lab Arg 2

be two complete argument labellings. Then LabArg2LabAsm( Lab Arg 1 ) =

LabArg2LabAsm( Lab Arg 2 ). Lemmas 18 and 19 imply that every complete assumption labelling is the translation of a unique complete argument labelling in terms of LabArg2LabAsm. Corollary 20. Let Lab Asm be a complete assumption labelling. Then there exists a unique complete argument labelling Lab Arg such that LabArg2LabAsm( Lab Arg ) = Lab Asm. From Theorems 15 and 17, Proposition 16, and Lemmas 18 and 19 it follows that there is a one-to-one correspondence between complete argument labellings and complete assumption labellings in terms of both LabArg2LabAsm and LabAsm2LabArg. Thus, when restricting LabArg2LabAsm and LabAsm2LabArg to complete argument and assumption labellings, they are bijective functions as well as the inverse of one another (see [24] for the analogous result about complete argument and assumption extensions). Corollary 21. Let L AsmComp be the set of all complete assumption labellings of L, R, A, ¯ and L ArgComp the set of all complete argument labellings of  Ar A B A , Att A B A . Let

• LabArg2LabAsm : L ArgComp → L AsmComp LabArg2LabAsm( Lab Arg ).

such

that

∀ Lab Arg ∈ L ArgComp : LabArg2LabAsm ( Lab Arg ) =

C. Schulz, F. Toni / International Journal of Approximate Reasoning 84 (2017) 110–149

127

• LabAsm2LabArg : L AsmComp → L ArgComp such that ∀ Lab Asm ∈ L AsmComp : LabAsm2LabArg ( Lab Asm) = LabAsm2LabArg( Lab Asm). LabArg2LabAsm and LabAsm2LabArg are bijective functions and each other’s inverse. 5.3. Grounded, preferred, ideal, and stable semantics Due to existing correspondence results between grounded, preferred, ideal, and stable argument labellings and extensions [25,32], argument and assumption extensions [23,20,24], and assumption extensions and labellings (see Section 3.3), the one-to-one correspondence between grounded, preferred, ideal, and stable assumption and argument labellings can be proven in a similar way as for complete assumption and argument labellings. Theorem 22 states the relationship between a given grounded/preferred/ideal/stable assumption labelling and a grounded/preferred/ideal/stable argument labelling in terms of LabAsm2LabArg. Theorem 22. Let Lab Asm be an assumption labelling. Lab Asm is a grounded/preferred/ ideal/stable assumption labelling if and only if LabAsm2LabArg( Lab Asm) is a grounded/ preferred/ideal/stable argument labelling. Theorem 23 states the relationship between a given grounded/preferred/ideal/stable argument labelling and a grounded/ preferred/ideal/stable assumption labelling in terms of LabArg2LabAsm. Theorem 23. Let Lab Arg be an argument labelling. If Lab Arg is a grounded/preferred/ ideal/stable argument labelling then

LabArg2LabAsm( Lab Arg ) is a grounded/preferred/ideal/ stable assumption labelling. Note that as for the complete semantics the converse of Theorem 23 does not hold. A counter-example is A B A 9 from Example 15 whose only grounded/preferred/ideal/stable assumption labelling is Lab Asm1 , which is the translation of the argument labelling Lab Arg 2 in terms of LabArg2LabAsm, but Lab Arg 2 is not a grounded/preferred/ideal/stable argument labelling. Due to the one-to-one correspondence between complete assumption and argument labellings (see Corollary 21), it is straightforward that there is also a one-to-one correspondence between grounded/preferred/ideal/stable argument and assumption labellings in terms of LabAsm2LabArg and LabArg2LabAsm. 5.4. Semi-stable semantics In contrast to the grounded, preferred, ideal, and stable semantics, semi-stable assumption and argument extensions are not in a one-to-one correspondence [24]. Since semi-stable assumption labellings correspond to semi-stable assumption extensions (Theorem 5) and semi-stable argument labellings to semi-stable argument extensions [25], it follows that there is no one-to-one correspondence between semi-stable assumption and argument labellings in terms of LabAsm2LabArg and LabArg2LabAsm. However, since semi-stable assumption and argument labellings are complete labellings, the translation of a semi-stable assumption labelling in terms of LabAsm2LabArg is of course a complete argument labelling and the translation of a semi-stable argument labelling in terms of LabArg2LabAsm is a complete assumption labelling. The following example illustrates an ABA framework where all semi-stable argument labellings are translated to semistable assumption labellings by LabArg2LabAsm, but not all semi-stable assumption labellings are translated to semistable argument labellings by LabAsm2LabArg. Example 16. Let A B A 10 be the ABA framework with:

L = {ρ , ψ, χ , x}, R = {x ← ψ, χ }, A = {ρ , ψ, χ }, ρ = ψ , ψ = ρ, χ = χ . A B A 10 has three complete assumption labellings: Lab Asm1 labels all assumptions as undec, and Lab Asm2 and Lab Asm3 are as illustrated in the ABA graphs in Fig. 11. Both Lab Asm2 and Lab Asm3 are semi-stable assumption labellings of A B A 10 . The corresponding AA framework of A B A 10 is shown in Fig. 12, along with two of its complete argument labellings Lab Arg 2 and Lab Arg 3 . The third complete argument labelling Lab Arg 1 labels all arguments as undec. Only Lab Arg 2 is a semi-stable argument labelling. Thus, LabArg2LabAsm translates all semi-stable argument labellings to semi-stable assumption labellings, namely LabArg2LabAsm( Lab Arg 2 ) = Lab Asm2 , but LabAsm2LabArg does not translate all semistable assumption labellings to semi-stable argument labellings since LabAsm2LabArg( Lab Asm3 ) = Lab Arg 3 .

128

C. Schulz, F. Toni / International Journal of Approximate Reasoning 84 (2017) 110–149

Fig. 11. The ABA graph of A B A 10 with two of its complete assumption labellings Lab Asm2 (left) and Lab Asm3 (right), which are both semi-stable assumption labellings (see Example 16).

Fig. 12. The AA graph of the corresponding AA framework of A B A 10 with two of its complete argument labellings Lab Arg 2 (left) and Lab Arg 3 (right). Only Lab Arg 2 is a semi-stable argument labelling (see Example 16).

Fig. 13. The ABA graph of A B A 11 with two of its complete assumption labellings Lab Asm2 (left) and Lab Asm3 (right). Only Lab Asm3 is a semi-stable assumption labelling (see Example 17).

The next example illustrates an ABA framework where all semi-stable assumption labellings are translated to semi-stable argument labellings by LabAsm2LabArg, but not all semi-stable argument labellings are translated to semi-stable assumption labellings by LabArg2LabAsm. Example 17. Let A B A 11 be the following ABA framework:

L = {ρ , ψ, χ , ω, x, w }, R = {x ← ψ, χ ; w ← ω; w ← ψ}, A = {ρ , ψ, χ , ω}, ρ = ψ , ψ = ρ , χ = χ , ω = w. A B A 11 has three complete assumption labellings: Lab Asm1 labels all assumptions as undec, and Lab Asm2 and Lab Asm3 are as illustrated in the ABA graphs in Fig. 13. Only Lab Asm3 is a semi-stable assumption labelling. The corresponding AA framework of A B A 11 is shown in Fig. 14, along with two of its complete argument labellings Lab Arg 2 and Lab Arg 3 . The third complete argument labelling Lab Arg 1 labels all arguments as undec. Both Lab Arg 2 and Lab Arg 3 are semi-stable argument labellings. Thus, LabAsm2LabArg translates all semi-stable assumption labellings to semi-stable argument labellings, namely LabAsm2LabArg( Lab Asm3 ) = Lab Arg 3 , but LabArg2LabAsm does not translate all semi-stable argument labellings to semi-stable assumption labellings since LabArg2LabAsm( Lab Arg 2 ) = Lab Asm2 . Note that A B A 10 and A B A 11 are special cases illustrating that semi-stable assumption and argument labellings do not correspond in general. However, there are also cases where semi-stable argument and assumption labellings correspond as demonstrated by the following example.

C. Schulz, F. Toni / International Journal of Approximate Reasoning 84 (2017) 110–149

129

Fig. 14. The AA graph of the corresponding AA framework of A B A 11 with two of its complete argument labellings Lab Arg 2 (top) and Lab Arg 3 (bottom), which are both semi-stable argument labellings (see Example 17).

Example 18. Let A B A 12 be the same as A B A 11 but with χ = x. Then Lab Asm1 and Lab Asm3 are complete assumption labellings as before, but in Lab Asm2 χ is labelled in rather than undec, so both Lab Asm2 and Lab Asm3 are semi-stable assumption labellings. The corresponding AA framework of A B A 12 has the same complete argument labellings Lab Arg 1 and Lab Arg 3 as he corresponding AA framework of A B A 11 , but in Lab Arg 2 the argument {χ }  χ is labelled in rather than undec. Thus, both Lab Arg 2 and Lab Arg 3 are semi-stable argument labellings, corresponding to the two semi-stable assumption labellings of A B A 12 in terms of LabAsm2LabArg and LabArg2LabAsm. 5.5. Admissible semantics We have shown in Theorem 1 that admissible assumption extensions and labellings are in a one-to-one correspondence. Furthermore, we know that admissible assumption and argument extensions correspond [23], but this correspondence is not one-to-one as for the complete, grounded, preferred, ideal, and stable semantics, but one-to-many as illustrated by the following example. Example 19. Let A B A 13 be the ABA framework with

L = {ρ , ψ, p }, R = { p ← ψ}, A = {ρ , ψ}, ρ = ψ , ψ = ρ. The admissible assumption extensions of A B A 13 are Asms1 = {}, Asms2 = {ρ } and Asms3 = {ψ}. The corresponding AA framework has three arguments, A 1 : {ρ }  ρ , A 2 : {ψ}  ψ , and A 3 : {ψ}  p, and four admissible argument extensions, Args1 = {}, Args2 = { A 1 }, Args3 = { A 2 }, and Args4 = { A 2 , A 3 }. Args3 and Args4 both correspond to the admissible assumption extension Asms3 in the sense that Asms3 is the set of all assumptions occurring in the premises of arguments in both Args3 and Args4 [23]. Conversely, only Args4 corresponds to Asms3 in the sense that it is the set of all arguments whose premises are contained in Asms3 [23].

130

C. Schulz, F. Toni / International Journal of Approximate Reasoning 84 (2017) 110–149

In addition, the correspondence between admissible argument extensions and labellings is one-to-many rather than oneto-one [25]. This implies that the correspondence between admissible assumption and argument labellings is one-to-many rather than one-to-one. Thus, only some of the correspondence results analogous to those for complete semantics hold for admissible semantics. Theorem 24. Let

Lab Asm

be an assumption labelling. If

Lab Asm

is an admissible assumption labelling then

LabAsm2LabArg( Lab Asm) is an admissible argument labelling. Example 20. Consider again A B A 13 from Example 19. A B A 13 has the same number of admissible assumption labellings and extensions, which correspond one-to-one:

• Lab Asm1 = {(ρ , undec), (ψ, undec)} corresponds to Asms1 ; • Lab Asm2 = {(ρ , in), (ψ, out)} corresponds to Asms2 ; and • Lab Asm3 = {(ρ , out), (ψ, in)} corresponds to Asms3 . In contrast, the corresponding AA framework of A B A 13 has eight admissible argument labellings, even though it has only four admissible argument extensions:

• Lab Arg 11 = {( A 1 , undec), ( A 2 , undec), ( A 3 , undec)} corresponds to Args1 ; • Lab Arg 21 = {( A 1 , in), ( A 2 , undec), ( A 3 , undec)}, Lab Arg 22 = {( A 1 , in), ( A 2 , undec), ( A 3 , out)}, Lab Arg 23 = {( A 1 , in), ( A 2 , out), ( A 3 , undec)}, and Lab Arg 24 = {( A 1 , in), ( A 2 , out), ( A 3 , out)} all correspond to Args2 ; • Lab Arg 31 = {( A 1 , undec), ( A 2 , in), ( A 3 , undec)}, and Lab Arg 32 = {( A 1 , out), ( A 2 , in), ( A 3 , undec)} both correspond to Args3 ; and • Lab Arg 41 = {( A 1 , out), ( A 2 , in), ( A 3 , in)} corresponds to Args4 . The translation of each admissible assumption labelling in terms of LabAsm2LabArg is an admissible argument labelling, i.e. LabAsm2LabArg( Lab Asm1 ) = Lab Arg 11 , LabAsm2LabArg( Lab Asm2 ) = Lab Arg 24 , and LabAsm2LabArg( Lab Asm3 ) = Lab Arg 41 . The following example shows that the converse of Theorem 24 does however not hold. Example 21. In A B A 13 from Example 20 LabAsm2LabArg({(ρ , in), (ψ, undec)}) = Lab Arg 21 , which is an admissible argument labelling, but {(ρ , in), (ψ, undec)} is not an admissible assumption labelling. It is furthermore not the case that every admissible argument labelling is the translation of some admissible assumption labelling in terms of LabAsm2LabArg (i.e. the analogous result of Proposition 16 for the admissible semantics does not hold). Example 22. Consider the admissible argument labelling Lab Arg 22 of A B A 13 (see Example 20). There exists no admissible assumption labelling such that Lab Arg 22 is the translation in terms of LabAsm2LabArg since the arguments A 2 and A 3 have different labels even though their premises are the same. Concerning LabArg2LabAsm, it is surprisingly not the case that the translation of every admissible argument labelling in terms of LabArg2LabAsm is an admissible assumption labelling (i.e. the analogous result of Theorem 17 for admissible semantics does not hold), as illustrated by the following example. Example 23. Consider the admissible argument labelling Lab Arg 31 of A B A 13 (see Example 20). LabArg2LabAsm( Lab Arg 31 ) = {(ρ , undec), (ψ, in)}, which is not an admissible assumption labelling. However, it holds that every admissible assumption labelling is the translation of some admissible argument labelling in terms of LabArg2LabAsm. Proposition 25. Let Lab Asm be an admissible assumption labelling. Then there exists an admissible argument labelling Lab Arg such that LabArg2LabAsm( Lab Arg ) = Lab Asm. As in the case of complete assumption and argument labellings, an admissible assumption labelling may also be the translation of some argument labelling in terms of LabArg2LabAsm which is not an admissible argument labelling. For

C. Schulz, F. Toni / International Journal of Approximate Reasoning 84 (2017) 110–149

131

example, LabArg2LabAsm({( A 1 , undec), ( A 2 , undec), ( A 3 , in)}) = Lab Asm2 , where Lab Asm2 is an admissible assumption labelling but {( A 1 , undec), ( A 2 , undec), ( A 3 , in)} is not an admissible argument labelling of A B A 13 (see Example 20). 5.5.1. Committed admissible argument labellings One of the reasons for the one-to-many correspondence between admissible assumption and argument labellings is the one-to-many correspondence between admissible argument extensions and labellings. This arises since admissible argument labellings make no restriction on arguments labelled undec, so any argument can be labelled undec in an admissible argument labelling. In contrast, admissible assumption labellings and extensions are in a one-to-one correspondence since admissible assumption labellings make restrictions on assumptions labelled undec (see Section 3.1). We now introduce a variant of admissible argument labellings, which follows the spirit of admissible assumption labellings by restricting undec arguments to arguments which are not attacked by any in labelled arguments. Definition 13. Let  Ar , Att  be an AA framework and let Lab Arg be an argument labelling of  Ar , Att . Lab Arg is a committed admissible argument labelling of  Ar , Att  if and only if for each argument A ∈ Ar it holds that:

• if Lab Arg ( A ) = in then for each B ∈ Ar attacking A, Lab Arg ( B ) = out; • if Lab Arg ( A ) = out then there exists some B ∈ Ar attacking A such that Lab Arg ( B ) = in; • if Lab Arg ( A ) = undec then there exists no B ∈ Ar attacking A such that Lab Arg ( B ) = in. From Definition 13 it follows directly that each committed admissible argument labelling is an admissible argument labelling. Corollary 26. Let  Ar , Att  be an AA framework and let Lab Arg be an argument labelling of  Ar , Att . If Lab Arg is a committed admissible argument labelling of  Ar , Att  then it is an admissible argument labelling of  Ar , Att , but not vice versa. Example 24. The AA framework  Ar A B A 13 , Att A B A 13  (see Examples 19 and 20) has four committed admissible argument labellings, namely Lab Arg 11 , Lab Arg 24 , Lab Arg 32 , and Lab Arg 41 . The other admissible argument labellings are not committed admissible since they violate the third condition in Definition 13. For example, Lab Arg 21 is not a committed admissible argument labelling since argument A 2 is labelled undec, but there exists an argument attacking A 2 which is labelled in, namely A 1 . Differently from admissible argument labellings, committed admissible argument labellings are in a one-to-one correspondence with admissible argument extensions. Theorem 27. Let  Ar , Att  be an AA framework. 1. Let Args ⊆ Ar be an admissible argument extension of  Ar , Att . Then Lab Arg with in( Lab Arg ) = Args, out( Lab Arg ) = Args+ , and undec( Lab Arg ) = Ar \ ( Args ∪ Args+ ) is a committed admissible argument labelling of  Ar , Att . 2. Let Lab Arg be a committed admissible argument labelling of  Ar , Att . Then Args = in( Lab Arg ) is an admissible argument extension of  Ar , Att  with Args+ = out( Lab Arg ), and Ar \ ( Args ∪ Args+ ) = undec( Lab Arg ). Note that the way arguments are labelled in, out, and undec in the first item of Theorem 27 mirrors the Ext2Lab operator in [25]. On the other hand, the second item of Theorem 27 extends the Lab2Ext operator from [25] as it not only defines an argument extension based on an argument labelling, but also the set of arguments attacked by the argument extension and the set of arguments which are neither contained in nor attacked by the argument extension. Note also that committed admissible argument labellings are different from other variations of the admissible semantics such as strongly admissible argument labellings (and extensions) [37,38], which require that an accepted argument is defended by accepted arguments other than itself, and related admissible argument extensions [16], which require that all accepted arguments are “relevant” for defending some accepted argument. Given this one-to-one correspondence between committed admissible argument labellings and admissible argument labellings, we now show that there is a “more refined” one-to-many correspondence between admissible assumption labellings and committed admissible argument labellings as compared to admissible argument labellings, i.e. some additional correspondence results hold. Firstly, the converse of Theorem 24 is satisfied for committed admissible argument labellings. Theorem 28. Let Lab Asm be an assumption labelling. Lab Asm is an admissible assumption labelling if and only if LabAsm2LabArg( Lab Asm) is a committed admissible argument labelling. Secondly, the translation of a committed admissible argument labelling in terms of LabArg2LabAsm is an admissible assumption labelling.

132

C. Schulz, F. Toni / International Journal of Approximate Reasoning 84 (2017) 110–149

Theorem 29. Let Lab Arg be an argument labelling. If Lab Arg is a committed admissible argument labelling then LabArg2LabAsm( Lab Arg ) is an admissible assumption labelling. Furthermore, Proposition 25 also holds for committed admissible argument labellings. Proposition 30. Let Lab Asm be an admissible assumption labelling. Then there exists a committed admissible argument labelling Lab Arg such that LabArg2LabAsm( Lab Arg ) = Lab Asm. The following example illustrates that due to the additional correspondence results, the one-to-many correspondence of admissible assumption labellings with committed admissible argument labellings is a “more refined” than with admissible argument labellings. Example 25. Consider again A B A 13 from Examples 19, 20, and 24. Lab Asm2 is the translation of only one committed admissible argument labelling in terms of LabArg2LabAsm, namely Lab Arg 24 , rather than of two different admissible argument labellings Lab Arg 23 and Lab Arg 24 . Furthermore, the translations of all committed admissible argument labellings in terms of LabArg2LabAsm are admissible assumption labellings. In contrast, the translations of the three admissible argument labellings Lab Arg 21 , Lab Arg 22 , and Lab Arg 31 in terms of LabArg2LabAsm are not admissible assumption labellings. The reason that despite the additional correspondence results there is no one-to-one correspondence between admissible assumption labellings and committed admissible argument labellings is that a committed admissible argument labelling may not be the translation of any admissible assumption labelling in terms of LabAsm2LabArg . For example, the committed admissible argument labelling Lab Arg 32 of A B A 13 is not the translation of any admissible assumption labelling in terms of LabAsm2LabArg (see Examples 19, 20, and 24). Note that it would be straightforward to define a new notion of admissible assumption labellings, which corresponds more closely to admissible argument labellings. This can be achieved by deleting the restriction on undec assumptions from the definition of admissible assumption labellings. However, we do not examine this possible variation further since we believe that the restriction on undec assumptions is intuitive: it seems reasonable that any assumption attacked by accepted assumptions cannot be accepted and should thus be rejected (out) rather than neither accepted nor rejected (undec). 6. Non-flat ABA frameworks So far, we only considered flat ABA frameworks. In general however, ABA frameworks may not be flat, for example the instance of ABA corresponding to auto-epistemic logic [39] is never flat [4]. For possibly non-flat ABA frameworks assumption extensions are defined more generally than for flat ABA frameworks: they are closed sets of assumptions and they are based on a more general notion of defence [4]. A set of assumption Asms ⊆ A

• is closed if and only if Asms = {α ∈ A | ∃ Asms ⊆ Asms : Asms  α }; • defends α ∈ A if and only if Asms attacks all closed sets of assumptions attacking α . Note that in flat ABA frameworks every set of assumptions is closed since in these frameworks assumptions do not occur as the head of inference rules and therefore the more general notion of defence coincides with the notion of defence introduced in Section 2. For flat ABA frameworks, the general definition of assumption extensions for possibly non-flat ABA frameworks (introduced in the following) thus coincides with the definitions given in Section 2. From here onwards, and if not specified otherwise, we assume as given a possibly non-flat ABA framework L, R, A, ¯. Furthermore, “defence” refers to the more general notion. 6.1. Admissible semantics We recall the definition of admissible assumption extensions for possibly non-flat ABA frameworks. A set of assumption Asms ⊆ A is an admissible assumption extension if and only if Asms is closed, conflict-free, and defends every α ∈ Asms. We first illustrate that admissible assumption labellings as introduced for flat ABA frameworks (Definition 2) do not correctly express the semantics of non-flat ABA frameworks. Example 26. Let A B A 14 be the non-flat ABA framework with:

L = {ρ , ψ, χ , p , x}, R = {ρ ← χ } ,

C. Schulz, F. Toni / International Journal of Approximate Reasoning 84 (2017) 110–149

133

A = {ρ , ψ, χ }, ρ = ψ , ψ = p, χ = x. According to Definition 2, A B A 14 has four admissible assumption labellings:

• • • •

Lab Asm1 = {(ρ , undec), (ψ, undec), (χ , undec)}, Lab Asm2 = {(ρ , out), (ψ, in), (χ , undec)}, Lab Asm3 = {(ρ , undec), (ψ, undec), (χ , in)}, and Lab Asm4 = {(ρ , out), (ψ, in), (χ , in)}.

However, A B A 14 has only two admissible assumption extensions (according to the definition for possibly non-flat ABA frameworks): Asms1 = {}, and Asms2 = {ψ}. Asms1 corresponds to Lab Asm1 and Asms2 to Lab Asm2 (in terms of Theorem 1). The corresponding sets of assumptions (in terms of Theorem 1) of Lab Asm3 and Lab Asm4 are Asms3 = {χ } and Asms4 = {ψ, χ }, respectively. Neither of them is an admissible assumption extension of A B A 14 , since neither of them is a closed set of assumptions. Thus, Lab Asm3 and Lab Asm4 should not be admissible assumption labellings of A B A 14 . As illustrated in Example 26, a reason that the definition of admissible assumption labellings of flat ABA frameworks does not correctly express the semantics of non-flat ABA frameworks is that the set of in labelled assumptions may not be closed. A straightforward way of revising the definition of admissible assumption labellings is thus to explicitly add the condition “in( Lab Asm) is a closed set of assumptions”. However, this condition expresses a restriction on the whole set of in labelled assumptions, rather than on the label of one assumption, as done by the three conditions of admissible assumption labellings. To adhere to the structure of the conditions of admissible assumption labellings, we instead add an additional restriction to the conditions of undec and out labelled assumptions, which ensures that an assumption can only be labelled undec or out if it is not derivable from the set of in labelled assumptions using the inference rules. To express this new restriction, we introduce the notion of a set of assumptions supporting an assumption. Definition 14. Let Asms ⊆ A and α ∈ A. Asms supports α if and only if there exists an argument Asms  α and Asms ⊆ Asms. Equivalently, we say that α is supported by Asms. The following definition extends Definition 2 to admissible assumption labellings of possibly non-flat ABA frameworks. Definition 15. Let Lab Asm be an assumption labelling. Lab Asm is an admissible assumption labelling if and only if for each assumption α ∈ A it holds that:

• if Lab Asm(α ) = in then for each closed set of assumptions Asms attacking α there exists some β ∈ Asms such that Lab Asm(β) = out; • if Lab Asm(α ) = out then there exists a closed set of assumptions Asms1 attacking α such that for all β ∈ Asms1 , Lab Asm(β) = in, and there exists no set of assumptions Asms2 supporting α such that for all γ ∈ Asms2 , Lab Asm(γ ) = in;

• if Lab Asm(α ) = undec then for each closed set of assumptions Asms1 attacking α there exists some β ∈ Asms1 such that Lab Asm(β) = in, and there exists no set of assumptions Asms2 supporting α such that for all γ ∈ Asms2 , Lab Asm(γ ) = in. According to the revised definition only Lab Asm1 and Lab Asm2 of A B A 14 from Example 26 are admissible assumption labellings. Lab Asm3 and Lab Asm4 are not admissible assumption labellings since ρ violates the new restriction on undec/out labelled assumptions as ρ is supported by {χ } and χ is labelled in. Note that we also incorporated the more general notion of defence into Definition 15 by only considering closed sets of assumptions attacking an assumption in question. Observation 31. Let Lab Asm be an assumption labelling of a flat ABA framework. Then Lab Asm is an admissible assumption labelling according to Definition 2 if and only if it is an admissible assumption labelling according to Definition 15. The following theorem states that Definition 15 correctly expresses the admissible semantics of possibly non-flat ABA frameworks, i.e. that there is a one-to-one correspondence between admissible assumption extensions and labellings of possibly non-flat ABA frameworks. Theorem 32. 1. Let Asms be an admissible assumption extension. Then Lab Asm with in( Lab Asm) = Asms, out( Lab Asm) = Asms+ and undec( Lab Asm) = A \ ( Asms ∪ Asms+ ) is an admissible assumption labelling.

134

C. Schulz, F. Toni / International Journal of Approximate Reasoning 84 (2017) 110–149

2. Let Lab Asm be an admissible assumption labelling. Then Asms = in( Lab Asm) is an admissible assumption extension with Asms+ = out( Lab Asm) and A \ ( Asms ∪ Asms+ ) = undec( Lab Asm). 6.2. Complete semantics We first recall the definition of complete assumption extensions for possibly non-flat ABA frameworks. A set of assumption Asms ⊆ A is a complete assumption extension if and only if Asms is closed, conflict-free, and consists of all assumptions it defends. For flat ABA frameworks, complete assumption labellings are defined as admissible assumption labellings satisfying an additional condition. Analogously, we define complete assumption labellings of possibly non-flat ABA frameworks. Definition 16. Let Lab Asm be an assumption labelling. Lab Asm is a complete assumption labelling if and only if Lab Asm is an admissible assumption labelling and for each assumption α ∈ A it holds that:

• if Lab Asm(α ) = undec then there exists a closed set of assumptions Asms3 attacking α such that for all δ ∈ Asms3 , Lab Asm(δ) = out. Analogously to the definition of admissible assumption labellings of possibly non-flat ABA frameworks, the additional condition of complete assumption labellings only takes into account attacking sets of assumptions which are closed. Without this restriction, the definition would yield different assumption labellings. Observation 33. Let Lab Asm be an assumption labelling of a flat ABA framework. Then Lab Asm is a complete assumption labelling according to Definition 3 if and only if it is a complete assumption labelling according to Definition 16. As intended, complete assumption labellings and extensions of possibly non-flat ABA frameworks are in one-to-one correspondence. Theorem 34. 1. Let Asms be a complete assumption extension. Then Lab Asm with in( Lab Asm) = Asms, out( Lab Asm) = Asms+ and undec( Lab Asm) = A \ ( Asms ∪ Asms+ ) is a complete assumption labelling. 2. Let Lab Asm be a complete assumption labelling. Then Asms = in( Lab Asm) is a complete assumption extension with Asms+ = out( Lab Asm) and A \ ( Asms ∪ Asms+ ) = undec( Lab Asm). For flat ABA frameworks we identified two equivalent variations to the definition of complete assumption labellings. One of them used the converse of each condition in the definition of a complete assumption labellings (see Lemma 3). Extending this alternative definition of complete assumption labellings of flat ABA frameworks with an additional condition ensuring that the set of in labelled assumptions is closed and considering only attacking sets of assumptions which are closed, makes it equivalent to the definition of complete assumption labellings for possibly non-flat ABA frameworks. Proposition 35. Let Lab Asm be an assumption labelling. The following statements are equivalent: 1. Lab Asm is a complete assumption labelling. 2. Lab Asm is such that for each α ∈ A it holds that: • if there exists a set of assumptions Asms supporting α such that for all β ∈ Asms, Lab Asm(β) = in, then Lab Asm(α ) = in; • if for each closed set of assumptions Asms attacking α there exists some β ∈ Asms such that Lab Asm(β) = out, then Lab Asm(α ) = in; • if there exists a closed set of assumptions Asms attacking α such that for all β ∈ Asms, Lab Asm(β) = in, then Lab Asm(α ) = out; • if for each closed set of assumptions Asms1 attacking α there exists some β ∈ Asms1 such that Lab Asm(β) = in, and there exists a closed set of assumptions Asms2 attacking α such that for all γ ∈ Asms2 , Lab Asm(γ ) = out, then Lab Asm(α ) = undec. Note that reversing the conditions in Definition 16 does not result in an equivalent definition of complete assumption labellings for possibly non-flat ABA frameworks. For example, the assumption labelling Lab Asm = {(ρ , undec), (ψ, in), (χ , in)} of A B A 14 (see Example 26) satisfies the converse of each condition in Definition 16: for both ψ and χ the converse of the first condition applies and is satisfied, and for ρ none of the converses of the three conditions applies, so ρ trivially satisfies the converse conditions. However, Lab Asm is not a complete assumption labelling of A B A 14 since A B A 14 has no complete assumption labellings.

C. Schulz, F. Toni / International Journal of Approximate Reasoning 84 (2017) 110–149

135

The other equivalent definition of complete assumption labellings for flat ABA frameworks we identified was the “if and only if” version of the first and second conditions of a complete assumption labelling for flat ABA frameworks (see Lemma 3). The analogue for complete assumption labellings of possibly non-flat ABA frameworks does however not result in an equivalent definition. That is, an assumption labelling satisfying the following conditions

• Lab Asm(α ) = in if and only if for each closed set of assumptions Asms attacking α there exists some β ∈ Asms such that Lab Asm(β) = out; • Lab Asm(α ) = out if and only if there exists a closed set of assumptions Asms1 attacking α such that for all β ∈ Asms1 , Lab Asm(β) = in, and there exists no set of assumptions Asms2 supporting α such that for all γ ∈ Asms2 , Lab Asm(γ ) = in; is not generally a complete assumption labelling of a possibly non-flat ABA framework since for instance Lab Asm = {(ρ , undec), (ψ, in), (χ , in)} of A B A 14 (see Example 26) satisfies both conditions, but Lab Asm is not a complete assumption labelling of A B A 14 . 6.3. Grounded, preferred, ideal, semi-stable, and stable semantics Originally, the grounded, preferred, and stable assumption extensions of possibly non-flat ABA frameworks were defined as specific admissible rather than complete assumption extensions. For flat ABA frameworks these two definitions are equivalent, bus as we will show in this section for non-flat ABA frameworks they are not. We first recall the definitions of grounded, preferred, and stable assumption extensions for possibly non-flat ABA frameworks [4]. A set of assumptions Asms ⊆ A is

• a grounded assumption extension if and only if Asms is the intersection of all complete assumption extensions; • a preferred assumption extension if and only if Asms is a maximal (w.r.t. ⊆) admissible assumption extension; • a stable assumption extension if and only if Asms is closed, conflict-free, and for all α ∈ A it holds that if α ∈ / Asms then Asms attacks

α.

Since ideal and semi-stable semantics have only been defined for flat ABA frameworks so far, we will investigate these semantics after dealing with the grounded, preferred, and stable semantics. 6.3.1. Grounded semantics The following example illustrates that for possibly non-flat ABA frameworks, the minimally complete assumption extensions do not generally coincide with the grounded assumption extensions. Example 27. Let A B A 15 be the non-flat ABA framework with:

L = {ρ , ψ, χ , ω, x}, R = {x ← ρ ; x ← ψ; χ ← }, A = {ρ , ψ, χ , ω}, ρ = ψ , ψ = ρ , χ = ω, ω = x. A B A 15 has two complete assumption extensions: Asms1 = {ρ , χ } and Asms2 = {ψ, χ }. Asms1 and Asms2 are both minimally complete, but the grounded assumption extension is Asms3 = {χ }. In order to express the grounded semantics of possibly non-flat ABA frameworks in terms of assumption labellings, the set of in labelled assumptions has to be the intersection of the sets of in labelled assumptions of all complete assumption labellings. Definition 17. Let Lab Asm be an assumption labelling. Lab Asm is a grounded assumption labelling if and only if for all it holds that:

α∈A

• Lab Asm(α ) = in if and only if for all complete assumption labellings Lab Asm , Lab Asm (α ) = in; • Lab Asm(α ) = out if and only if there exists a closed set of assumptions Asms attacking α such that for all β ∈ Asms, Lab Asm(β) = in. The second condition ensures the one-to-one correspondence between grounded assumption labellings and extensions of possibly non-flat ABA frameworks.

136

C. Schulz, F. Toni / International Journal of Approximate Reasoning 84 (2017) 110–149

Theorem 36. 1. Let Asms be a grounded assumption extension. Then Lab Asm with in( Lab Asm) = Asms, out( Lab Asm) = Asms+ and undec( Lab Asm) = A \ ( Asms ∪ Asms+ ) is a grounded assumption labelling. 2. Let Lab Asm be a grounded assumption labelling. Then Asms = in( Lab Asm) is a grounded assumption extension with Asms+ = out( Lab Asm) and A \ ( Asms ∪ Asms+ ) = undec( Lab Asm). Based on the correspondence between grounded assumption labellings and extensions of possibly non-flat ABA frameworks and results from [4], we prove that for flat ABA frameworks Definition 17 is equivalent to the definition of grounded assumption labellings for flat ABA frameworks. Proposition 37. Let Lab Asm be an assumption labelling of a flat ABA framework. Then Lab Asm is a grounded assumption labelling according to Definition 4 if and only if it is a grounded assumption labelling according to Definition 17. 6.3.2. Preferred semantics The non-flat ABA framework A B A 14 from Example 26 illustrates that maximally complete assumption extensions do not generally coincide with preferred assumption extensions: A B A 14 has no complete assumption extensions, but {ψ} is its preferred assumption extension. We thus define preferred assumption labellings of possibly non-flat ABA frameworks as admissible, rather than complete, assumption labellings with a maximal set of in labelled assumptions. Definition 18. Let Lab Asm be an assumption labelling. Lab Asm is a preferred assumption labelling if and only if Lab Asm is an admissible assumption labelling and in( Lab Asm) is maximal (w.r.t. ⊆) among all admissible assumption labellings. Since preferred assumption labellings for flat ABA frameworks can be equivalently defined as admissible assumption labellings with a maximal set of in labelled assumptions (see Proposition 7) and since for flat ABA frameworks Definition 15 coincides with Definition 2, it follows that for flat ABA frameworks Definition 18 coincides with the definition of preferred assumption labellings of flat ABA frameworks. Proposition 38. Let Lab Asm be an assumption labelling of a flat ABA framework. Then Lab Asm is a preferred assumption labelling according to Definition 4 if and only if it is a preferred assumption labelling of according to Definition 18. As desired, preferred assumption labellings correctly express the preferred semantics of possibly non-flat ABA frameworks. Theorem 39. 1. Let Asms be a preferred assumption extension. Then Lab Asm with in( Lab Asm) = Asms, out( Lab Asm) = Asms+ and undec( Lab Asm) = A \ ( Asms ∪ Asms+ ) is a preferred assumption labelling. 2. Let Lab Asm be a preferred assumption labelling. Then Asms = in( Lab Asm) is a preferred assumption extension with Asms+ = out( Lab Asm) and A \ ( Asms ∪ Asms+ ) = undec( Lab Asm). 6.3.3. Stable semantics Even though stable assumption extensions of possibly non-flat ABA frameworks are not defined as specific admissible or complete assumption extensions, it was shown in [4] that stable assumption extensions are always complete assumption extensions. Therefore, we define stable assumption labellings of possibly non-flat ABA frameworks in the same way as for flat ABA frameworks, i.e. as complete assumption labellings which label no assumptions as undec. Definition 19. Let Lab Asm be an assumption labelling. Lab Asm is a stable assumption labelling if and only if Lab Asm is a complete assumption labelling and undec( Lab Asm) = ∅. From Observation 33 and Definition 19 it follows straightaway that for flat ABA frameworks Definition 19 is equivalent to the definition of stable assumption labellings of flat ABA frameworks. Observation 40. Let Lab Asm be an assumption labelling of a flat ABA framework. Then Lab Asm is a stable assumption labelling according to Definition 4 if and only if it is a stable assumption labelling according to Definition 19. Furthermore, there is a one-to-one correspondence between stable assumption labellings and extensions of possibly non-flat ABA frameworks.

C. Schulz, F. Toni / International Journal of Approximate Reasoning 84 (2017) 110–149

137

Theorem 41. 1. Let Asms be a stable assumption extension. Then Lab Asm with in( Lab Asm) = Asms, out( Lab Asm) = Asms+ and undec( Lab Asm) = A \ ( Asms ∪ Asms+ ) is a stable assumption labelling. 2. Let Lab Asm be a stable assumption labelling. Then Asms = in( Lab Asm) is a stable assumption extension with Asms+ = out( Lab Asm) and A \ ( Asms ∪ Asms+ ) = undec( Lab Asm). 6.3.4. Ideal semantics Since the ideal semantics has so far only been defined in the context of flat ABA frameworks, we define ideal assumption extensions of possibly non-flat ABA frameworks, following the original definition for flat ABA frameworks where ideal assumption extensions are defined as specific admissible rather than complete assumption extensions [23]. Definition 20. A set of assumptions Asms ⊆ A is an ideal assumption extension if and only if Asms is a maximal (w.r.t. ⊆) admissible assumption extension satisfying that for all preferred assumption extensions Asms , Asms ⊆ Asms . Just as for preferred assumption extensions, ideal assumption extensions of possibly non-flat ABA frameworks do not generally coincide with maximally complete assumption extensions which are a subset of each preferred assumption extension. We thus define ideal assumption labellings of possibly non-flat ABA frameworks in terms of admissible rather than complete assumption labellings. Definition 21. Let Lab Asm be an assumption labelling. Lab Asm is an ideal assumption labelling if and only if Lab Asm is an admissible assumption labelling and in( Lab Asm) is maximal (w.r.t. ⊆) among all admissible assumption labellings satisfying that for all preferred assumption labellings Lab Asm , in( Lab Asm) ⊆ in( Lab Asm ). From Proposition 7 and Observation 31 it follows that for flat ABA frameworks Definition 21 coincides with the definition of ideal assumption labellings of flat ABA frameworks. Proposition 42. Let Lab Asm be an assumption labelling of a flat ABA framework. Then Lab Asm is an ideal assumption labelling according to Definition 4 if and only if it is an ideal assumption labelling according to Definition 21. Furthermore, as desired there is a one-to-one correspondence between ideal assumption extensions and labellings of possibly non-flat ABA frameworks. Theorem 43. 1. Let Asms be an ideal assumption extension. Then Lab Asm with in( Lab Asm) = Asms, out( Lab Asm) = Asms+ and undec( Lab Asm) = A \ ( Asms ∪ Asms+ ) is an ideal assumption labelling. 2. Let Lab Asm be an ideal assumption labelling. Then Asms = in( Lab Asm) is an ideal assumption extension with Asms+ = out( Lab Asm) and A \ ( Asms ∪ Asms+ ) = undec( Lab Asm). 6.3.5. Semi-stable semantics Just like the ideal semantics, the semi-stable semantics has so far only been defined for flat ABA frameworks. Semi-stable assumption extensions are originally defined as specific complete assumption extensions, but for flat ABA frameworks they can be equivalently defined as specific admissible assumption extensions. For non-flat ABA frameworks this is not the case. Example 28. Let A B A 16 be the non-flat ABA framework with:

L = {ρ , ψ, χ , ω, p }, R = { p ← ρ ; p ← χ ; p ← ψ; ψ ← ρ , χ }, A = {ρ , ψ, χ , ω}, ρ = ψ , ψ = p, χ = ψ , ω = χ . The only complete assumption extension of A B A 16 is Asms1 = {}, and thus Asms1 ∪ Asms+ 1 is maximal among all complete assumption extensions. In contrast, there are three admissible assumption extensions: Asms1 , Asms2 = {ρ }, and Asms3 = {χ }. Among these, Asms3 ∪ Asms+ 3 is maximal. One of the defining properties of semi-stable assumption extensions of flat ABA frameworks is that they are preferred assumption extensions [24]. To retain this property, we define semi-stable assumption extensions and labellings of possibly non-flat ABA frameworks in terms of admissible rather than complete assumption extensions and labellings.

138

C. Schulz, F. Toni / International Journal of Approximate Reasoning 84 (2017) 110–149

Definition 22. A set of assumptions Asms ⊆ A is a semi-stable assumption extension if and only if Asms is an admissible assumption extension and for all admissible assumption extensions Asms , Asms ∪ Asms+ ⊂ Asms ∪ Asms + . Definition 23. Let Lab Asm be an assumption labelling. Lab Asm is a semi-stable assumption labelling if and only if Lab Asm is an admissible assumption labelling and undec( Lab Asm) is minimal (w.r.t. ⊆) among all admissible assumption labellings. By Proposition 7, for flat ABA frameworks Definition 23 coincides with the definition of semi-stable assumption labellings of flat ABA frameworks. Proposition 44. Let Lab Asm be an assumption labelling of a flat ABA framework. Then Lab Asm is a semi-stable assumption labelling according to Definition 4 if and only if it is a semi-stable assumption labelling according to Definition 23. As desired, there is a one-to-one correspondence between semi-stable assumption extensions and labellings of possibly non-flat ABA frameworks. Theorem 45. 1. Let Asms be a semi-stable assumption extension. Then Lab Asm with in( Lab Asm) = Asms, out( Lab Asm) = Asms+ and undec( Lab Asm) = A \ ( Asms ∪ Asms+ ) is a semi-stable assumption labelling. 2. Let Lab Asm be a semi-stable assumption labelling. Then Asms = in( Lab Asm) is a semi-stable assumption extension with Asms+ = out( Lab Asm) and A \ ( Asms ∪ Asms+ ) = undec( Lab Asm). Finally, we prove that semi-stable assumption labellings of possibly non-flat ABA frameworks satisfy the property we desired, namely that they are preferred assumption labellings. Proposition 46. Let Lab Asm be a semi-stable assumption labelling. Then Lab Asm is a preferred assumption labelling. 7. Conclusion We studied assumption labellings of flat as well as possibly non-flat ABA frameworks for the admissible, grounded, complete, preferred, ideal, semi-stable, and stable semantics and proved that there is a one-to-one correspondence with the respective assumption extensions. Even though some of the definitions of assumption labellings for possibly non-flat ABA frameworks are (considerably) different from the definitions of assumption labellings for flat ABA frameworks, in particular those for the grounded, preferred, ideal, and semi-stable semantics, we proved that applying the definitions of assumption labellings for possibly non-flat ABA frameworks to flat ABA frameworks yields the correct assumption labellings (as defined for flat ABA frameworks). Furthermore, we showed that assumption labellings of flat ABA frameworks can be equivalently defined based on attacks between all sets of assumptions and attacks between argument-supporting sets of assumptions. Using the latter notion, we demonstrated how argument-supporting sets of assumptions and attacks between them can be naturally represented as a graph, which expresses sufficient information to determine the assumption labellings of a flat ABA framework. We also investigated the relationship of assumption labellings of flat ABA frameworks and argument labellings of their corresponding AA frameworks and found that grounded, complete, preferred, ideal, and stable assumption and argument labellings are in a one-to-one correspondence, whereas semi-stable assumption labellings and extensions do not generally correspond. Admissible assumption and argument labellings are in a one-to-many correspondence. Based on this finding, we defined a variant of admissible argument labellings, which we call committed admissible argument labellings and which mirrors admissible assumption labellings by requiring that arguments which are attacked by in labelled arguments are labelled out, rather than undec. We proved that admissible assumption labellings and committed admissible argument labellings are in a “more refined” one-to-many correspondence as compared to admissible argument labellings, i.e. an admissible assumption labelling may correspond to fewer committed admissible argument labellings than admissible argument labellings. In contrast to admissible argument labellings, committed admissible argument labellings are furthermore in one-to-one correspondence with admissible argument extensions. In this paper, we did not focus on any particular instance of ABA. For example, logic programs are special instances of flat ABA frameworks [4], so the results presented here for flat ABA frameworks also apply to the instance of ABA which corresponds to logic programs. Caminada and Schulz [29] show that, conversely, a specific fragment of (flat) ABA frameworks is as an instance of logic programs, where the models of the logic program and the assumption labellings of the ABA framework coincide. Furthermore, Dung [3] as well as Caminada et al. [40] study logic programs as an instance of AA frameworks, proving that the models of the logic program and the argument labellings of the AA framework correspond. In Section 5, we investigated (flat) ABA as a direct instance of AA (as for example in [23,24]) and analysed the (non-)correspondence between assumption and argument labellings in more detail than previously done for assumption and argument extensions [23,24], without restricting to any fragment or instance of flat ABA.

C. Schulz, F. Toni / International Journal of Approximate Reasoning 84 (2017) 110–149

139

Since the stable semantics of ABA, both in terms of assumption and argument extensions, corresponds to the stable model semantics for logic programs [4,22], ABA has proven useful to argumentatively explain the truth value (true or false) of literals with respect to stable models of logic programs [22]. Other logic programming semantics such as 3-valued stable models [41], well-founded models [42], and regular models [43] also include a third truth value (undefined) for literals. Since these 3-valued logic programming models correspond to assumption labellings [28,29], assumption labellings could be used to argumentatively explain the truth value of literals with respect to 3-valued logic programming models. Furthermore, assumption labellings may prove useful for explaining inconsistencies in logic programs under the stable model semantics since some inconsistency scenarios can be attributed to the “undefined” truth value of literals [44], which corresponds to undec assumptions in ABA frameworks [22,29]. The ABA graphs introduced here are fundamentally different from previous graphical representations of ABA frameworks, notably the attack relationship graphs in [4] and the argument graphs in [36]. As logic programs are special instances of ABA frameworks, ABA graphs can also be used to represent logic programs [28]. For logic programs, various other graphical representations have been proposed, which can be compared to ABA graphs of ABA frameworks instantiated by logic programs. Dimopoulos and Torres [45] introduce minimal attack graphs of logic programs, which are very similar to ABA graphs when used on ABA frameworks which are instantiated by a logic program. The difference is that ABA graphs may have more nodes since all argument-supporting sets of assumptions form nodes, rather than only minimal ones. Costantini [46] introduces Extended Dependency Graphs to represent a special type of logic programs. The graphs have one node for every rule, holding the rule’s head atom. The representation is thus rather different from ABA graphs, even when considering ABA frameworks which are instantiated by logic programs. Graphical representations of logic programs have also been used to compute the semantics of logic programs [47,48]. In future work, we will study the suitability of ABA graphs for computing assumption labellings of ABA frameworks. In future work we are planning to investigate the role of distinguishing between rejected and undecided assumptions in dispute derivations [34]. Currently, dispute derivations focus on determining whether an assumption is accepted with respect to some extension. If no dispute derivation can be constructed, the assumption in question is deemed non-accepted. We will explore whether assumption labellings can help to further differentiate between rejected and undecided assumptions in dispute derivations. Furthermore, ABA frameworks have recently been extended to incorporate preferences [49,50]. Future work will reveal whether assumption labellings can be directly applied to these extended ABA frameworks. Another recent development regarding ABA frameworks is the definition of semantics for ABA frameworks in terms of argument graphs rather than argument extensions [36]. It will be interesting to investigate if argument or assumption labellings can be directly applied to argument graphs or if a completely new labelling approach is needed. Argument labellings have inspired efficient algorithms for computing the semantics of an AA framework [51–53]. Assumption labellings may thus prove useful for algorithms to compute ABA semantics in terms of assumptions, which is another line of future research. Appendix A. Proofs Proof of Theorem 1 1. First note that Asms ∩ Asms+ = ∅ since Asms does not attack itself. Thus each α ∈ A is either contained in in( Lab Asm), in out( Lab Asm), or in undec( Lab Asm). • Let Lab Asm(α ) = in. Then α ∈ Asms, so Asms defends α , i.e. for all sets of assumptions Asms1 attacking α there exists some β ∈ Asms1 such that Asms attacks β . Thus, β ∈ Asms+ and consequently Lab Asm(β) = out. • Let Lab Asm(α ) = out. Then α ∈ Asms+ , so Asms attacks α . Since Asms = in( Lab Asm), there exists a set of assumptions Asms1 attacking α such that for all β ∈ Asms1 , Lab Asm(β) = in. • Let Lab Asm(α ) = undec. Then α ∈ / Asms and α ∈ / Asms+ , so α is not attacked and not defended by Asms. Since α is not attacked by Asms, for each set of assumptions Asms1 attacking α there exists some β ∈ Asms1 such that β∈ / Asms, and thus Lab Asm(β) = in. 2. We first prove that in( Lab Asm) is an admissible assumption extension. • in( Lab Asm) is conflict-free: Assume in( Lab Asm) is not conflict-free. Then in( Lab Asm) attacks some α ∈ in( Lab Asm). By Definition 2, for each set of assumptions Asms1 attacking α there exists some β ∈ Asms1 such that Lab Asm(β) = out. Hence, in( Lab Asm) contains some β such that Lab Asm(β) = out. Contradiction. • in( Lab Asm) defends all α ∈ in( Lab Asm): Let α ∈ in( Lab Asm). Then by Definition 2, for each set of assumptions Asms1 attacking α there exists some β ∈ Asms1 such that Lab Asm(β) = out. Furthermore, for each such β there exists a set of assumptions Asms2 attacking β such that for all γ ∈ Asms2 , Lab Asm(γ ) = in so Asms2 ⊆ in( Lab Asm). Hence, in( Lab Asm) attacks all sets of assumptions attacking α . Asms+ = {α ∈ A | Asms attacks α } = {α ∈ A | in( Lab Asm) attacks α } = {α ∈ A | α ∈ out( Lab Asm)} = out( Lab Asm). A \ ( Asms ∪ Asms+ ) = {α ∈ A | α ∈ / in( Lab Asm), α ∈ / out( Lab Asm)} = {α ∈ A | α ∈ undec( Lab Asm)} = undec( Lab Asm).

140

C. Schulz, F. Toni / International Journal of Approximate Reasoning 84 (2017) 110–149

Proof of Lemma 2 Since each set of assumptions attacking α contains some β such that Lab Asm(β) = out, Lab Asm(α ) = out. If Lab Asm(α ) = in, then ∀γ ∈ α : Lab Asm(γ ) = out and therefore Lab Asm = Lab Asm, so trivially Lab Asm is an admissible assumption labelling. If Lab Asm(α ) = undec, then ∀γ ∈ α : Lab Asm(γ ) = undec or Lab Asm(γ ) = out. Furthermore, α ∈ / α since each set of assumptions attacking α contains some β such that Lab Asm(β) = out, so if α ∈ α then ∃ Asms attacking α such that Lab Asm(α ) = out (since all ∀δ ∈ Asms : δ = α → Lab Asm(δ) = in), which is a contradiction. Therefore, Lab Asm is an assumption labelling.

• Let Lab Asm ( ) = in. If Lab Asm( ) = in, then for each set of assumption Asms1 attacking  there exists some η ∈ Asms1 such that Lab Asm(η) = out and therefore Lab Asm (η) = out. If Lab Asm( ) = in then  = α , so for each set of assumptions Asms2 attacking α there exists some β ∈ Asms such that Lab Asm(β) = out and thus Lab Asm (β) = out. • Let Lab Asm ( ) = out. If Lab Asm( ) = out then there exists a set of assumptions Asms1 attacking  such that for all η ∈ Asms1 , Lab Asm(η) = in and therefore Lab Asm (η) = in. If Lab Asm( ) = out then  ∈ α , so there exists a set of assumptions Asms2 attacking  such that ∀δ ∈ Asms2 with δ = α it holds that Lab Asm(δ) = in and thus Lab Asm (δ) = in. Since Lab Asm (α ) = in the set of assumptions Asms2 attacking  is such that for all η ∈ Asms3 , Lab Asm (η) = in. • Let Lab Asm ( ) = undec. Then Lab Asm( ) = undec. Thus, for each set of assumptions Asms1 attacking  there exists some η ∈ Asms1 such that Lab Asm(η) = in. Since  ∈ / α , for each such set of assumptions Asms1 attacking  either α ∈ / Asms1 or there exists some κ ∈ Asms1 such that κ = α and Lab Asm(κ ) = in. In the first case η = α , so Lab Asm (η) = in. In the second case, Lab Asm (κ ) = in. Thus, for each set of assumptions Asms1 attacking  there exists some λ ∈ Asms1 such that Lab Asm (λ) = in. Proof of Theorem 3 Equivalence of first and second item:

• First item implies second item: Let Lab Asm be a complete assumption labelling. Then clearly the “only if” part of both conditions of the second item are satisfied since they are the same as the conditions in Definition 3. To prove that the “if” part of the conditions in the second item holds: – Let α be an assumption such that for each set of assumptions Asms attacking α there exists some β ∈ Asms such that Lab Asm(β) = out. Then Lab Asm(α ) = out because there exists no set of assumptions Asms1 attacking α such that for all β ∈ Asms1 , Lab Asm(β) = in. Furthermore, Lab Asm(α ) = undec because there exists no set of assumptions Asms2 attacking α such that for all γ ∈ Asms2 , Lab Asm(γ ) = out. Hence, Lab Asm(α ) = in. – Let α be an assumption such that there exists a set of assumptions Asms attacking α such that for all β ∈ Asms, Lab Asm(β) = in. Then Lab Asm(α ) = in because not for each set of assumptions Asms1 attacking α there exists some β ∈ Asms1 such that Lab Asm(β) = out. Furthermore, Lab Asm(α ) = undec because not for each set of assumptions Asms2 attacking α there exists some γ ∈ Asms2 such that Lab Asm(γ ) = in. Hence, Lab Asm(α ) = out. • Second item implies first item: Let Lab Asm be such that the second item holds. We prove that Lab Asm is a complete assumption labelling. Clearly the first two conditions of complete assumption labellings are satisfied since they are the same as the “only if” part of the conditions in the second item. To prove the third condition of complete assumption labellings, let Lab Asm(α ) = undec. From the first condition of the second item we know that not for each set of assumptions Asms1 attacking α there exists some β ∈ Asms1 such that Lab Asm(β) = out, so there exists a set of assumptions Asms2 attacking α such that for all γ ∈ Asms2 , Lab Asm(γ ) = out. From the second condition of the second item we know that there exists no set of assumptions Asms3 attacking α such that for all δ ∈ Asms3 , Lab Asm(δ) = in, so for each set of assumptions Asms4 attacking α there exists some  ∈ Asms4 such that Lab Asm( ) = in. Equivalence of second and third item:

• Second item implies third item: Let Lab Asm be such that the second item holds. Then clearly the first two conditions of the third item are satisfied since they are the same as the “if” part of the second item. To prove the third condition of the third item, let α be such that for each set of assumptions Asms1 attacking α there exists some β ∈ Asms1 such that Lab Asm(β) = in, and there exists a set of assumptions Asms2 attacking α such that for all γ ∈ Asms2 , Lab Asm(γ ) = out. Then Lab Asm(α ) = in because not for each set of assumptions Asms3 attacking α there exists some δ ∈ Asms3 such that Lab Asm(δ) = out, and Lab Asm(α ) = out because there exists no set of assumptions Asms4 attacking α such that for all  ∈ Asms4 , Lab Asm( ) = in. Hence, Lab Asm(α ) = undec. • Third item implies second item: Assume that Lab Asm is such that the third item holds. Then clearly the “if” part of both conditions in the second item are satisfied since they the same as the conditions in the third item. To prove that the “only if” parts of the conditions in the second item are satisfied, first note that for every α ∈ A exactly one of the “if” parts of the three conditions in the third item is satisfied. Thus, if Lab Asm(α ) = in the “if” part of the second and third condition in the third item are not satisfied. It follows that the “if” part of the first condition is satisfied, so for each set of assumptions Asms1 attacking α there exists some β ∈ Asms1 such that Lab Asm(β) = out.

C. Schulz, F. Toni / International Journal of Approximate Reasoning 84 (2017) 110–149

141

Analogously, if Lab Asm(α ) = out only the “if” part of the second condition in the third item applies, so there exists a set of assumptions Asms2 attacking α such that for all β ∈ Asms2 , Lab Asm(β) = in. Proof of Theorem 4 1. Since Asms is a complete assumption extension it is by definition also an admissible assumption extension. By Theorem 1, Lab Asm is an admissible assumption labelling. It remains to prove that the additional condition of complete assumption labellings is satisfied. Let Lab Asm(α ) = undec. Then α ∈ / Asms and α ∈ / Asms+ , so α is not attacked and not defended by Asms. Since α is not defended by Asms, there exists a set of assumption Asms1 attacking α such that Asms1 is not attacked by Asms. Thus, for all γ ∈ Asms1 it holds that γ ∈ / Asms+ . Consequently, Lab Asm(γ ) = out. 2. Since Lab Asm is a complete assumption labelling it is by Definition 3 also an admissible assumption labelling. Thus, by Theorem 1 Asms is an admissible assumption extension with Asms+ = out( Lab Asm) and A \ ( Asms ∪ Asms+ ) = undec( Lab Asm). It remains to prove that all assumptions defended by Asms are contained in Asms. Let α be defended by Asms and thus by in( Lab Asm). Then for each set of assumptions Asms1 attacking α , in( Lab Asm) attacks Asms1 . Thus, for each such Asms1 there exists some β ∈ Asms1 which is attacked by in( Lab Asm), and therefore Lab Asm(β) = out. Since this holds for each Asms1 attacking α , Lab Asm(α ) = in. Proof of Theorem 5 1. Let Asms be a 1) grounded 2) preferred 3) ideal 4) semi-stable 5) stable assumption extension. By definition Asms is a complete assumption extension. Furthermore, for all complete assumption extensions Asms it holds that 1) Asms ⊂ Asms 2) Asms ⊃ Asms 3) if for all preferred assumption extensions Asms

it holds that Asms ⊆ Asms

then Asms ⊃ Asms 4) Asms ∪ Asms + ⊃ Asms ∪ Asms+ 5) Asms ∪ Asms+ = A. By Theorem 4 Lab Asm is a complete assumption labelling. Furthermore, from the above and Theorem 4, for all complete assumption labellings Lab Asm it holds that 1) in( Lab Asm ) ⊂ in( Lab Asm) 2) in( Lab Asm ) ⊃ in( Lab Asm) 3) if for all preferred assumption labellings Lab Asm

it holds that in( Lab Asm ) ⊆ in( Lab Asm

) then in( Lab Asm ) ⊃ in( Lab Asm) 4) in( Lab Asm ) ∪ out( Lab Asm ) ⊃ in( Lab Asm) ∪ out( Lab Asm), and consequently undec( Lab Asm ) ⊂ undec( Lab Asm) 5) in( Lab Asm) ∪ out( Lab Asm) = A, and consequently undec( Lab Asm) = ∅. Therefore, Lab Asm is a 1) grounded 2) preferred 3) ideal 4) semi-stable 5) stable assumption labelling. 2. Let Lab Asm be a 1) grounded 2) preferred 3) ideal 4) semi-stable 5) stable assumption labelling. By definition Lab Asm is a complete assumption labelling. Furthermore, for all complete assumption labellings Lab Asm it holds that 1) in( Lab Asm ) ⊂ in( Lab Asm) 2) in( Lab Asm ) ⊃ in( Lab Asm) 3) if for all preferred assumption labellings Lab Asm

it holds that in( Lab Asm ) ⊆ in( Lab Asm

) then in( Lab Asm ) ⊃ in( Lab Asm) 4) undec( Lab Asm ) ⊂ undec( Lab Asm), or equivalently in( Lab Asm ) ∪ out( Lab Asm ) ⊃ in( Lab Asm) ∪ out( Lab Asm) 5) undec( Lab Asm) = ∅, or equivalently in( Lab Asm) ∪ out( Lab Asm) = A. By Theorem 4 Asms = in( Lab Asm) is a complete assumption extension with Asms+ = out( Lab Asm) and A \ ( Asms ∪ Asms+ ) = undec( Lab Asm). Furthermore, from the above and by Theorem 4, for all complete assumption extensions Asms it holds that 1) Asms ⊂ Asms 2) Asms ⊃ Asms 3) if for all preferred assumption extensions Asms

it holds that Asms ⊆ Asms

then Asms ⊃ Asms 4) Asms ∪ Asms + ⊃ Asms ∪ Asms+ 5) Asms ∪ Asms+ = A. Therefore, Asms is a 1) grounded 2) preferred 3) ideal 4) semi-stable 5) stable assumption extension. Proof of Proposition 7

• Preferred: Follows from the one-to-one correspondence between complete assumption labellings and extensions (Theorem 4) and between admissible assumption labellings and extensions (Theorem 1) together with Theorem 8 in [24].

• Ideal: Follows from the one-to-one correspondence between complete assumption labellings and extensions (Theorem 4) and between admissible assumption labellings and extensions (Theorem 1) together with Theorem 10 in [24].

• Semi-stable: From left to right: Let Lab Asm be a semi-stable assumption labelling, i.e. a complete assumption labelling such that undec( Lab Asm) is minimal among all complete assumption labellings. By definition, Lab Asm is an admissible assumption labelling. Assume undec( Lab Asm) is not minimal among all admissible assumptions labellings, i.e. ∃ Lab Asm with undec( Lab Asm ) ⊂ undec( Lab Asm) and Lab Asm is an admissible assumption labelling but not a complete assumption labelling. Thus Lab Asm satisfies Definition 2 but not Definition 3, so ∃α ∈ undec( Lab Asm ) such that for all sets of assumptions Asms attacking α there exists some β ∈ Asms such that Lab Asm (β) = out. By Lemma 2, Lab Asm

with in( Lab Asm

) = in( Lab Asm ) ∪ {α }, out( Lab Asm

) = out( Lab Asm ) ∪ α , and undec( Lab Asm

) = undec( Lab Asm ) \ ({α } ∪ α ) is an admissible assumption labelling. Clearly, undec( Lab Asm

) ⊂ undec( Lab Asm ), so undec( Lab Asm ) is not minimal among all admissible assumption labellings. Contradiction. From right to left: Let Lab Asm be an admissible assumption labelling such that undec( Lab Asm) is minimal (w.r.t. ⊆) among all admissible assumption labellings. Assume that Lab Asm is not a complete assumption labelling. By the same reasoning as above, ∃α ∈ undec( Lab Asm) such that for all sets of assumptions Asms attacking α there exists some β ∈ Asms such that Lab Asm(β) = out. It follows that there exists an admissible assumption labelling Lab Asm

with undec( Lab Asm

) ⊂ undec( Lab Asm). Contradiction. Thus, Lab Asm is a complete assumption labelling. Furthermore,

142

C. Schulz, F. Toni / International Journal of Approximate Reasoning 84 (2017) 110–149

since for all admissible assumption labellings Lab Asm , undec( Lab Asm ) ⊂ undec( Lab Asm) and since every complete assumption labelling is an admissible assumption labelling, it follows that for all complete assumption labellings Lab Asm , undec( Lab Asm ) ⊂ undec( Lab Asm). Thus, undec( Lab Asm) is minimal (w.r.t. ⊆) among all complete assumption labellings. • Stable: From left to right: Let Lab Asm be a complete assumption labelling such that undec( Lab Asm) = ∅. By definition, Lab Asm is admissible. From right to left: Let Lab Asm be an admissible assumption labelling such that undec( Lab Asm) = ∅. Then Lab Asm is also a complete assumption labelling since the conditions for in and out assumptions are the same for admissible and complete assumption labellings (see Definitions 2 and 3). Proof of Lemma 8 Follows from Theorem 1 and Theorem 4.2 in [33]. Proof of Proposition 9

• From left to right: Let Lab Asm be a complete assumption labelling. By definition, Lab Asm is an admissible assumption labelling and by Lemma 8 an admissible assumption labelling with respect to argument-supporting sets. It remains to prove that the additional condition of complete assumption labellings with respect to argument-supporting sets is satisfied. Let Lab Asm(α ) = undec. Then there exists a set of assumptions Asms attacking α such that for all β ∈ Asms, Lab Asm(β) = out. Thus, there exists an argument Asms  α such that Asms ⊆ Asms. Therefore, Asms is an argument-supporting set of assumptions attacking α such that for all γ ∈ Asms , Lab Asm(γ ) = out. • From right to left: Let Lab Asm be a complete assumption labelling with respect to argument-supporting sets. By definition, Lab Asm is and admissible assumption labelling with respect to argument-supporting sets and by Lemma 8 an admissible assumption labelling. It remains to prove that the additional condition of complete assumption labellings is satisfied. Let Lab Asm(α ) = undec. Then there exists an argument-supporting set of assumptions Asms attacking α such that for all γ ∈ Asms, Lab Asm(γ ) = out. Proof of Proposition 11 Let there be an infinite sequence of edges (α1 , α2 ), (α2 , α3 ), . . . in Gatt . Then there exists a set of assumptions Asmsα1 attacking α2 such that α1 ∈ Asmsα1 , a set of assumptions Asmsα2 attacking α3 such that α2 ∈ Asmsα2 , and so on. Thus, in G there exists an edge from Asmsα1 to {α2 } as well as to every other set of assumptions containing α2 , in particular an edge to Asmsα2 . Furthermore, there is an edge from Asmsα2 to {α3 } as well as every set of assumptions containing α3 , and so on. Thus, there is an infinite sequence of edges ( Asmsα1 , Asmsα2 ), ( Asmsα2 , Asmsα3 ), . . . in G . Example 13 proves that the converse does not hold. Proof of Lemma 13 Note first that LabAsm2LabArg is clearly a function.

• Injective: We prove that no two different assumption labellings Lab Asm1 and Lab Asm2 are mapped to the same argument labelling by LabAsm2LabArg. Let Lab Asm1 = Lab Asm2 . Thus, ∃α ∈ A such that Lab Asm1 (α ) = Lab Asm2 (α ). If α ∈ in( Lab Asm1 ) then α ∈ / in( Lab Asm2 ), so {α }  α ∈ in(LabAsm2LabArg( Lab Asm1 )) but {α }  α ∈ / in(LabAsm2LabArg( Lab Asm2 )). Analogous results are reached assuming that α ∈ out( Lab Asm1 ) and that α ∈ undec( Lab Asm1 ). Thus, LabAsm2LabArg( Lab Asm1 ) = LabAsm2LabArg( Lab Asm2 ). • Not generally surjective: The following ABA framework illustrates that there may be some Lab Arg ∈ L Arg such that there exists no Lab Asm ∈ L Asm with Lab Arg = LabAsm2LabArg( Lab Asm): L = {r , ρ }, R = {r ← }, A = {ρ }, ρ = r. There are three possible assumption labellings: Lab Asm1 = {(ρ , in)}, Lab Asm2 = {(ρ , out)}, and Lab Asm3 = {(ρ , undec)}. The corresponding AA framework has two arguments: Ar A B A = { A 1 : {ρ }  ρ , A 2 : {}  r }. In the translations of all three assumption labellings in terms of LabAsm2LabArg, A 2 is labelled in. Thus, for instance for the argument labelling {( A 1 , in), ( A 2 , out)} there exists no Lab Asm such that LabAsm2LabArg( Lab Asm) = {( A 1 , in), ( A 2 , out)}. Proof of Lemma 14 Note first that LabArg2LabAsm is clearly a function.

• Surjective: We prove that for every Lab Asm ∈ L Asm there exists some Lab Arg ∈ L Arg such that LabArg2LabAsm( Lab Arg ) = Lab Asm. Let Lab Asm ∈ L Asm . Furthermore, let Lab Arg be an argument labelling which

C. Schulz, F. Toni / International Journal of Approximate Reasoning 84 (2017) 110–149

143

satisfies that for all α ∈ A, {α }  α ∈ in( Lab Arg ) if α ∈ in( Lab Asm), {α }  α ∈ out( Lab Arg ) if α ∈ out( Lab Asm), and {α }  α ∈ undec( Lab Arg ) if α ∈ undec( Lab Asm). Then LabArg2LabAsm( Lab Arg ) = Lab Asm. Clearly Lab Arg ∈ L Arg . • Not generally injective: Consider the ABA framework from the proof of Lemma 13 and the two argument labellings Lab Arg 1 = {( A 1 , in), ( A 2 , out)} and Lab Arg 2 = {( A 1 , in), ( A 2 , in)}. Then LabArg2LabAsm( Lab Arg 1 ) = LabArg2LabAsm( Lab Arg 2 ) = {(ρ , in)}. Lemma 47. Let Asms ⊆ A and Args = { Asms  s ∈ Ar A B A | Asms ⊆ Asms}.

• Args+ = { Asms  s ∈ Ar A B A | ∃α ∈ Asms : α ∈ Asms+ }; • Ar A B A \ ( Args ∪ Args+ ) = { Asms  s ∈ Ar A B A | Asms  Asms, α ∈ Asms : α ∈ Asms+ }. Proof. We prove both statements:

• Args+ = { Asms  s ∈ Ar A B A | Args attacks Asms  s} = { Asms  s ∈ Ar A B A | ∃α ∈ Asms : ∃ Asms

 α ∈ Args} = { Asms  s ∈ Ar A B A | ∃α ∈ Asms : ∃ Asms

 α and Asms

⊆ Asms} = { Asms  s ∈ Ar A B A | ∃α ∈ Asms : Asms attacks α } = { Asms  s ∈ Ar A B A | ∃α ∈ Asms : α ∈ Asms+ }. • Ar A B A \ ( Args ∪ Args+ ) = { Asms  s ∈ Ar A B A | Asms  s ∈ / Args, Asms  s ∈ / Args+ } = { Asms  s ∈ Ar A B A | Asms 

+ Asms, α ∈ Asms : α ∈ Asms }. Lemma 48. Let Args ⊆ Ar A B A and let Asms = {α ∈ A | ∃ Asms : α ∈ Asms and Asms  s ∈ Args}. Then

• Asms+ = {α ∈ A | {α }  α ∈ Args+ }; • A \ ( Asms ∪ Asms+ ) = {α ∈ A | {α }  α ∈ / Args, {α }  α ∈ / Args+ }. Proof. We prove both statements:

• Asms+ = {α ∈ A | Asms attacks α } = {α ∈ A | ∃ Asms  α : Asms ⊆ Asms} = {α ∈ A | ∃ Asms  α ∈ Args} = {α ∈ A | Args attacks {α }  α } = {α ∈ A | {α }  α ∈ Args+ }. • A \ ( Asms ∪ Asms+ ) = {α ∈ A | α ∈ / Asms, α ∈ / Asms+ } = {α ∈ A |  Asms : α ∈ Asms and Asms  s ∈ Args, {α }  α ∈ / Args+ } + = { α ∈ A | {α }  α ∈ / Args, {α }  α ∈ / Args }. Proof of Theorem 15

• From left to right: Let Lab Asm be a complete assumption labelling. Firstly note that for all Asms  s ∈ Ar A B A exactly one of the three conditions in the definition of LabAsm2LabArg applies, so all Asms  s are in exactly one of in( Lab Arg ), out( Lab Arg ), or undec( Lab Arg ). By Theorem 4: Asms = in( Lab Asm) is a complete assumption extension with Asms+ = out( Lab Asm) and A \ ( Asms ∪ Asms+ ) = undec( Lab Asm). By Theorem 6.1 in [24]: Args = { Asms  s | Asms ⊆ in( Lab Asm)} is a complete argument extension. By Lemma 47: Args+ = { Asms  s | ∃α ∈ Asms : α ∈ out( Lab Asm)} and Ar A B A \ ( Args ∪ Args+ ) = { Asms  s | Asms  in( Lab Asm), α ∈ Asms : α ∈ out( Lab Asm)} = { Asms  s | ∃α ∈ Asms : α ∈ undec( Lab Asm), Asms ∩ out( Lab Asm) = ∅}. By Theorem 10 in [25]: Lab Arg with in( Lab Arg ) = Args, out( Lab Arg ) = Args+ , undec( Lab Arg ) = Ar A B A \ ( Args ∪ Args+ ) is a complete argument labelling. • From right to left: Let Lab Arg = LabAsm2LabArg( Lab Asm) be a complete argument labelling where Lab Asm is an assumption labelling. Since LabAsm2LabArg is injective by Lemma 13, Lab Asm is unique. By Theorems 9 and 11 in [25]: Args = in( Lab Arg ) = { Asms  s | Asms ⊆ in( Lab Asm)} is a complete argument extension with Args+ = out( Lab Arg ) = { Asms  s | ∃α ∈ Asms : α ∈ out( Lab Asm)} and Ar A B A \ ( Args ∪ Args+ ) = undec( Lab Arg ) = { Asms  s | ∃α ∈ Asms : α ∈ undec( Lab Asm), Asms ∩ out( Lab Asm) = ∅}. By Theorem 6.1 in [24]: Asms = {α ∈ A | ∃ Asms : α ∈ Asms and Asms  s ∈ Args} = in( Lab Asm) is a complete assumption extension. By Lemma 48: Asms+ = {α | {α }  α ∈ Args+ } = {α | α ∈ out( Lab Asm)} = out( Lab Asm) and A \ ( Asms ∪ Asms+ ) = { α | {α }  α ∈ / Args, {α }  α ∈ / Args+ } = {α | {α }  α ∈ undec( Lab Arg )} = {α | α ∈ undec( Lab Asm)} = undec( Lab Asm). By Theorem 4: Lab Asm is a complete assumption labelling. Proof of Proposition 16 By Theorem 9 in [25], Args = in( Lab Arg ) is a complete argument extension.

144

C. Schulz, F. Toni / International Journal of Approximate Reasoning 84 (2017) 110–149

By Theorem 11 in [25], Args+ = out( Lab Arg ) and Ar A B A \ ( Args ∪ Args+ ) = undec( Lab Arg ). By Theorem 6.1 in [24], Asms = {α | ∃ Asms : α ∈ Asms , Asms  s ∈ Args} is a complete assumption extension. From Theorem 6.1 and Proposition 1 in [24] it also follows that Args = { Asms  s | Asms ⊆ Asms}. By Lemma 47, Args+ = { Asms  s | ∃α ∈ Asms : α ∈ Asms+ }, and Ar A B A \ ( Args ∪ Args+ ) = { Asms  s | Asms  Asms, α ∈ Asms : α ∈ Asms+ }. By Theorem 4, Lab Asm with in( Lab Asm) = Asms, out( Lab Asm) = Asms+ , and undec( Lab Asm) = A \ ( Asms ∪ Asms+ ) is a complete assumption labelling. It follows that Args = { Asms  s | Asms ⊆ in( Lab Asm)} = in( Lab Arg ). Furthermore, Args+ = { Asms  s | ∃α ∈ Asms : α ∈ out( Lab Asm)} = out( Lab Arg ), and Ar A B A \ ( Args ∪ Args+ ) = { Asms  s | Asms  in( Lab Asm), α ∈ Asms : α ∈ out( Lab Asm)} = { Asms  s | ∃α ∈ Asms : α ∈ undec( Lab Asm), Asms ∩ out( Lab Asm) = ∅} = undec( Lab Arg ). Thus, LabAsm2LabArg( Lab Asm) = Lab Arg. Since LabAsm2LabArg is injective by Lemma 13, Lab Asm is unique. Proof of Theorem 17 By Theorems 9 and 11 in [25], Args = in( Lab Arg ) is a complete argument extension with Args+ = out( Lab Arg ) and Ar A B A \ ( Args ∪ Args+ ) = undec( Lab Arg ). By Theorem 6.1 in [24], Asms = {α | ∃ Asms : α ∈ Asms and Asms  s ∈ in( Lab Arg )} is a complete assumption extension. By Lemma 48, Asms+ = {α | {α }  α ∈ Args+ } = {α | {α }  α ∈ out( Lab Arg )} and A \ ( Asms ∪ Asms+ ) = {α | {α }  α ∈ / Args, {α }  α ∈ / Args+ } = {α | {α }  α ∈ undec( Lab Arg )}. Since for an argument Asms  s ∈ in( Lab Arg ) it holds that all attackers are labelled out, it follows that ∀α ∈ Asms : all attackers of {α }  α are labelled out, so by the definition of complete argument labellings {α }  α ∈ in( Lab Arg ). Thus, Asms = {α | {α }  α ∈ in( Lab Arg )}. By Theorem 4, Lab Asm with in( Lab Asm) = Asms, out( Lab Asm) = Asms+ and undec( Lab Asm) = A \ ( Asms ∪ Asms+ ) is a complete assumption labelling. Proof of Lemma 18 Let Lab Arg = LabAsm2LabArg( Lab Asm), so by Theorem 15 Lab Arg is a complete argument labelling. Now let Lab Asm = LabArg2LabAsm( Lab Arg ), so in( Lab Asm ) = {α | {α } ⊆ in( Lab Asm)} = in( Lab Asm), out( Lab Asm ) = {α | α ∈ out( Lab Asm)} = out( Lab Asm), undec( Lab Asm ) = {α | α ∈ undec( Lab Asm), {α } ∩ out( Lab Asm) = ∅} = undec( Lab Asm). Thus, Lab Asm = Lab Asm , so there exists a complete argument labelling Lab Arg such that LabArg2LabAsm( Lab Arg ) = Lab Asm. Proof of Lemma 19 Let LabArg2LabAsm( Lab Arg 1 ) = Lab Asm1 and LabArg2LabAsm( Lab Arg 2 ) = Lab Asm2 . Assume by contradiction that Lab Asm1 = Lab Asm2 . Since Lab Arg 1 = Lab Arg 2 , ∃ Asms1  s1 ∈ Ar A B A such that Lab Arg 1 ( Asms1  s1 ) = Lab Arg 2 ( Asms1  s 1 ).

• Let Asms1  s1 ∈ in( Lab Arg 1 ), so Asms1  s1 ∈ / in( Lab Arg 2 ). Then there exists some Asms2  α attacking Asms1  s1 where α ∈ Asms1 and Asms2  α ∈ / out( Lab Arg 2 ). However, Asms2  α ∈ out( Lab Arg 1 ) since all attackers of Asms1  s1 are labelled out by Lab Arg 1 . Thus, {α }  α ∈ / in( Lab Arg 2 ) but {α }  α ∈ in( Lab Arg 1 ), so α ∈ in( Lab Asm1 ) but α ∈in( / Lab Asm2 ). Contradiction. • Let Asms1  s1 ∈ out( Lab Arg 1 ), so Asms1  s1 ∈ / out( Lab Arg 2 ). Then there exists some Asms2  α attacking Asms1  s1 where α ∈ Asms1 and Asms2  α ∈ in( Lab Arg 1 ). However, Asms2  α ∈ / in( Lab Arg 2 ) since no attacker of Asms1  s1 is labelled in by Lab Arg 2 . Thus, {α }  α ∈ out( Lab Arg 1 ) but {α }  α ∈ / out( Lab Arg 2 ), so LabArg2LabAsm that α ∈ out( Lab Asm1 ) but α ∈/ out( Lab Asm2 ). Contradiction. • Let Asms1  s1 ∈ undec( Lab Arg 1 ), so Asms1  s1 ∈ / undec( Lab Arg 2 ). Then either for all Asms2  α attacking Asms1  s1 where α ∈ Asms1 it holds that Asms2  α ∈ out( Lab Arg 2 ) or there exists some Asms3  β attacking Asms1  s1 where β ∈ Asms1 and Asms3  β ∈ in( Lab Arg 2 ). In the first case for all {α }  α , {α }  α ∈ in( Lab Arg 2 ) but some {α }  α ∈ undec( Lab Arg1 ) since there exists an attacker Asms2  α of Asms1  s1 such that Asms2  α ∈/ out( Lab Arg1 ). It follows that α ∈ undec( Lab Asm1 ) but α ∈ / undec( Lab Asm2 ). Contradiction. In the second case, {β}  β ∈ out( Lab Arg 2 ) but {β}  β ∈ / out( Lab Arg 1 ) since no attacker of Asms1  s1 is labelled in by Lab Arg 1 . Thus, β ∈ out( Lab Asm2 ) but β∈ / out( Lab Asm1 ). Contradiction. Proof of Theorem 22 Analogous to the proof of Theorem 15 but using Theorem 5 instead of Theorem 4, Theorem 6.2/6.3/6.4/6.5 in [24] instead of Theorem 6.1 in [24], the analogues of Theorems 10 and 11 in [25] for the grounded/preferred/stable semantics (only informally given in [25]) instead of Theorems 9, 10, and 11 in [25], and Theorem 3.7 in [32] for the ideal semantics instead of Theorems 9, 10, and 11 in [25].

C. Schulz, F. Toni / International Journal of Approximate Reasoning 84 (2017) 110–149

145

Proof of Theorem 23 Analogous to the proof of Theorem 17 but using Theorem 5 instead of Theorem 4, Theorem 6.2/6.3/6.4/6.5 in [24] instead of Theorem 6.1 in [24], the analogues of Theorems 9 and 11 in [25] for the grounded/preferred/stable semantics (only informally given in [25]) instead of Theorems 9 and 11 in [25], and Theorem 3.7 in [32] for the ideal semantics instead of Theorems 9 and 11 in [25]. Proof of Theorem 24 Analogous to the “from left to right” part of the proof of Theorem 15 but using Theorem 1 instead of Theorem 4, Theorem 2.2 in [23] instead of Theorem 6.1 in [24], and Theorem 21 in [25] instead of Theorem 10 in [25]. Proof of Proposition 25 Analogous to the proof of Lemma 18 but using Theorem 24 instead of Theorem 15. Proof of Theorem 27 1. First note that Args ∩ Args+ = ∅ since Args does not attack itself. Thus each A ∈ Ar is either contained in in( Lab Arg ), out( Lab Arg ), or undec( Lab Arg ), so Lab Arg is an argument labelling. We prove that Lab Arg satisfies Definition 13. • Let Lab Arg ( A ) = in. Then A ∈ Args. Thus, all attackers B of A are attacked by some C ∈ Args, so B ∈ Args+ . Consequently, for each attacker B of A, Lab Arg ( B ) = out. • Let Lab Arg ( A ) = out. Then A ∈ Args+ . Thus, A is attacked by some B ∈ Args, and therefore there exists some B attacking A such that Lab Arg ( B ) = in. • Let Lab Arg ( A ) = undec. Then A ∈ / Args+ . Thus, A is not attacked by any B ∈ Args and consequently there exists no B attacking A such that Lab Arg ( B ) = in. 2. We prove that in( Lab Arg ) is an admissible argument extension. • in( Lab Arg ) is conflict-free: Assume in( Lab Arg ) is not conflict-free. Then there exist A , B ∈ in( Lab Arg ) such that A attacks B, so B is attacked by an argument which is not labelled out. Contradiction. • All arguments in in( Lab Arg ) are defended by in( Lab Arg ): Let A ∈ in( Lab Arg ). Then for each attacker B of A, Lab Arg ( B ) = out and therefore for each B there exists an attacker C such that Lab Arg (C ) = in. Thus, each attacker of A is attacked by in( Lab Arg ), i.e. in( Lab Arg ) defends A. Args+ = { A | Args attacks A } = { A | in( Lab Arg ) attacks A } = { A | A ∈ out( Lab Arg )} = out( Lab Arg ). Ar \ ( Args ∪ Args+ ) = { A | A ∈ / Args, A ∈ / Args+ } = {A | A ∈ / in( Lab Arg ), A ∈ / out( Lab Arg )} = { A | A ∈ undec( Lab Arg )} = undec( Lab Arg ). Proof of Theorem 28 Analogous to the proof of Theorem 15, but using Theorem 1 instead of Theorem 4, Theorem 27 instead of Theorems 10 and 11 in [25], and Theorem 2.2 in [23] instead of Theorem 6.1 in [24]. Proof of Theorem 29 Analogous to the proof of Theorem 17 but using Theorem 27 instead of Theorems 9 and 11 in [25], Theorem 2.2 in [23] instead of Theorem 6.1 in [24], and Theorem 1 instead of Theorem 4. Proof of Proposition 30 Analogous to the proof of Lemma 18 but using Theorem 28 instead of Theorem 15. Proof of Theorem 32 1. First note that Asms ∩ Asms+ = ∅ since Asms does not attack itself. Thus each α ∈ A is either contained in in( Lab Asm), in out( Lab Asm), or in undec( Lab Asm). We prove that Lab Asm satisfies Definition 15. • Let Lab Asm(α ) = in. Then α ∈ Asms, so Asms defends α , i.e. for all closed sets of assumptions Asms1 attacking α there exists some β ∈ Asms1 such that Asms attacks β . Thus, β ∈ Asms+ and consequently Lab Asm(β) = out. Therefore, for each closed set of assumptions Asms1 attacking α there exists some β ∈ Asms1 such that Lab Asm(β) = out.

146

C. Schulz, F. Toni / International Journal of Approximate Reasoning 84 (2017) 110–149

• Let Lab Asm(α ) = out. Then α ∈ Asms+ , so Asms attacks α . Since Asms = in( Lab Asm) and since Asms is a closed set of assumptions, there exists a closed set of assumptions Asms1 attacking α such that for all β ∈ Asms1 , Lab Asm(β) = in. Furthermore, since Asms is a closed set of assumptions, for all δ supported by Asms it holds that δ ∈ Asms. Since α ∈ Asms+ and since Asms ∩ Asms+ = ∅, it follows that α ∈/ Asms and therefore α is not supported by Asms. Thus there exists no set of assumptions Asms2 supporting α such that for all γ ∈ Asms2 , Lab Asm(γ ) = in. • Let Lab Asm(α ) = undec. Then α ∈ / Asms and α ∈ / Asms+ , so α is not attacked and not defended by Asms. Since α is not attacked by Asms, for each closed set of assumptions Asms1 attacking α there exists some β ∈ Asms1 such that β ∈ / Asms, and thus Lab Asm(β) = in. Furthermore, since Asms is a closed set of assumptions, it follows from the same reasoning as in the previous item that there exists no set of assumptions Asms2 supporting α such that for all γ ∈ Asms2 , Lab Asm(γ ) = in. 2. We first prove that in( Lab Asm) is an admissible assumption extension. • in( Lab Asm) is closed: Assume in( Lab Asm) is not closed. Then ∃α ∈ / in( Lab Asm) such that in( Lab Asm) supports α . Thus, Lab Asm(α ) = out or Lab Asm(α ) = undec. Contradiction since in either case there exists no set of assumptions Asms1 supporting α such that for all γ ∈ Asms1 , Lab Asm(γ ) = in. • in( Lab Asm) is conflict-free: Assume in( Lab Asm) is not conflict-free. Then in( Lab Asm) attacks some α ∈ in( Lab Asm). By Definition 15, for each closed set of assumptions Asms1 attacking α there exists some β ∈ Asms1 such that Lab Asm(β) = out. Since in( Lab Asm) is a closed set of assumptions, there exists some β ∈ in( Lab Asm) such that Lab Asm(β) = out. Contradiction. • in( Lab Asm) defends all α ∈ in( Lab Asm): Let α ∈ in( Lab Asm). Then by Definition 15, for each closed set of assumptions Asms1 attacking α there exists some β ∈ Asms1 such that Lab Asm(β) = out. Furthermore, for each such β there exists a closed set of assumptions Asms2 attacking β such that for all γ ∈ Asms2 , Lab Asm(γ ) = in so Asms2 ⊆ in( Lab Asm). Hence, in( Lab Asm) attacks all closed sets of assumptions attacking α . Asms+ = out( Lab Asm) and A \ ( Asms ∪ Asms+ ) = undec( Lab Asm) as in the proof of Theorem 1. Proof of Theorem 34 1. Since Asms is a complete assumption extension it is by definition also an admissible assumption extension [4]. By Theorem 32, Lab Asm is an admissible assumption labelling. It remains to prove that the additional condition of complete assumption labellings is satisfied. Let Lab Asm(α ) = undec. Then α ∈ / Asms and α ∈ / Asms+ , so α is not attacked and not defended by Asms. Since α is not defended by Asms, there exists a closed set of assumption Asms1 attacking α such that Asms1 is not attacked by Asms. Thus, for all γ ∈ Asms1 it holds that γ ∈ / Asms+ . Consequently, Lab Asm(γ ) = out. 2. Since Lab Asm is a complete assumption labelling it is by Definition 16 also an admissible assumption labelling. Thus, by Theorem 32 Asms is an admissible assumption extension with Asms+ = out( Lab Asm) and A \ ( Asms ∪ Asms+ ) = undec( Lab Asm). It remains to prove that all assumptions defended by Asms are contained in Asms. Let α be defended by Asms and thus by in( Lab Asm). Then for each closed set of assumptions Asms1 attacking α , in( Lab Asm) attacks Asms1 . Thus, for each such Asms1 there exists some β ∈ Asms1 which is attacked by in( Lab Asm), and therefore Lab Asm(β) = out. Since this holds for each Asms1 attacking α , Lab Asm(α ) = in. Proof of Proposition 35 First item implies second item:

• Let α be such that there exists a set of assumptions Asms supporting α such that for all β ∈ Asms, Lab Asm(β) = in. If Lab Asm(α ) = out or Lab Asm(α ) = undec then the second or third, respectively, condition of complete assumption labellings is violated. Thus Lab Asm(α ) = in since it satisfies the first condition. • Let α be such that for each closed set of assumptions Asms attacking α there exists some β ∈ Asms such that Lab Asm(β) = out. If Lab Asm(α ) = out or Lab Asm(α ) = undec then the second or third, respectively, condition of complete assumption labellings is violated. Thus Lab Asm(α ) = in since it satisfies the first condition. • Let α be such that there exists a closed set of assumptions Asms attacking α such that for all β ∈ Asms, Lab Asm(β) = in. If Lab Asm(α ) = in or Lab Asm(α ) = undec then the first or third, respectively, condition of complete assumption labellings is violated. Thus Lab Asm(α ) = out since it satisfies the second condition. • Let α be such that for each closed set of assumptions Asms1 attacking α there exists some β ∈ Asms1 such that Lab Asm(β) = in, and there exists a closed set of assumptions Asms2 attacking α such that for all γ ∈ Asms2 , Lab Asm(γ ) = out. If Lab Asm(α ) = in or Lab Asm(α ) = out then the first or second, respectively, condition of complete assumption labellings is violated. Thus Lab Asm(α ) = undec since it satisfies the third condition. Second item implies first item.

• Let Lab Asm(α ) = in. Then for each closed set of assumptions Asms1 attacking α there exists some β ∈ Asms1 such that Lab Asm(β) = in. Furthermore, it either holds that there exists a closed set of assumptions Asms2 attacking α such that

C. Schulz, F. Toni / International Journal of Approximate Reasoning 84 (2017) 110–149

147

for all γ ∈ Asms2 , Lab Asm(γ ) = in, (contradiction) or that for each closed set of assumptions Asms3 attacking α there exists some δ ∈ Asms3 such that Lab Asm(δ) = out. Thus, the second part of the or-statement applies. • Let Lab Asm(α ) = out. Then there exists no set of assumptions Asms1 supporting α such that for all β ∈ Asms1 , Lab Asm(β) = in. Furthermore, there exists a closed set of assumptions Asms2 attacking α such that for all γ ∈ Asms2 , Lab Asm(γ ) = out. Furthermore, it either holds that there exists a closed set of assumptions Asms3 attacking α such that for all δ ∈ Asms3 , Lab Asm(δ) = in, or that for each closed set of assumptions Asms4 attacking α there exists some  ∈ Asms4 such that Lab Asm( ) = out (contradiction). Thus, the first part of the or-statement applies. • Let Lab Asm(α ) = undec. Then there exists no set of assumptions Asms1 supporting α such that for all β ∈ Asms1 , Lab Asm(β) = in. Furthermore, there exists a closed set of assumptions Asms2 attacking α such that for all γ ∈ Asms2 , Lab Asm(γ ) = out. Furthermore, for each closed set of assumptions Asms3 attacking α there exists some δ ∈ Asms3 such that Lab Asm(δ) = in. Proof of Theorem 36 1. First note that since Asms is the intersection of all complete assumption labellings, which are all conflict-free, it follows that Asms ∩ Asms+ = ∅ and thus each α ∈ A is either contained in in( Lab Asm), out( Lab Asm), or undec( Lab Asm), so Lab Asm is an assumption labelling. Furthermore, note that grounded assumption extensions are always closed, even though this is not explicitly required in their definition. Since the grounded assumption extension is a subset of every complete assumption extension, any assumption α supported by the grounded assumption extension is also supported by each complete assumption extension. Since complete assumption extensions are closed, α is thus in each complete assumption extension and consequently part of the grounded assumption extension. We prove that Lab Asm satisfies Definition 17. • From left to right: Let Lab Asm(α ) = in. Then α ∈ Asms. Therefore, for all complete assumption extensions Asms , α ∈ Asms . By Theorem 34, for each Asms it holds that Lab Asm with in( Lab Asm ) = Asms , out( Lab Asm ) = Asms + , and undec( Lab Asm ) = A \ ( Asms ∪ Asms + ) is a complete assumption labelling and there are no other complete assumption labellings. Thus, for all complete assumption labellings Lab Asm , Lab Asm (α ) = in. From right to left: Let α be such that for all complete assumption labellings Lab Asm , Lab Asm (α ) = in. Then by Theorem 34, for each Lab Asm it holds that Asms = in( Lab Asm ) is a complete assumption extension and there are no other complete assumption extensions. Thus, for all complete assumption extensions Asms , α ∈ Asms . Therefore, α ∈ Asms and thus Lab Asm(α ) = in. • From left to right: Let Lab Asm(α ) = out. Then α ∈ Asms+ . Thus, α is attacked by Asms. Assume Asms is not closed. Then ∃β ∈ / Asms and Asms1 ⊆ Asms such that Asms1  β . Since for all complete assumption extensions Asms it holds that Asms ⊆ Asms and Asms is closed, it follows that β ∈ Asms . Thus, β ∈ Asms, which is a contradiction. Thus, Asms is closed, so α is attacked by a closed set of assumptions all labelled in by Lab Asm. From right to left: Let α be such that there exists a closed set of assumptions Asms1 attacking α such that for all β ∈ Asms1 , Lab Asm(β) = in. Then Asms1 ⊆ Asms, so α is attacked by Asms. Therefore, α ∈ Asms+ and thus Lab Asm(α ) = out. 2. Since Lab Asm is a grounded assumption labelling, it holds that for all α ∈ A: Lab Asm(α ) = in if and only if for all complete assumption labellings Lab Asm , Lab Asm (α ) = in. By Theorem 34, for each Lab Asm it holds that Asms = in( Lab Asm ) with Asms + = out( Lab Asm ) and A \ ( Asms ∪ Asms + ) = undec( Lab Asm ) is a complete assumption extension and there are no other complete assumption extensions. Thus, for all α ∈ A: α ∈ Asms if and only if for all complete assumption extensions Asms , α ∈ Asms . Therefore, Asms is the intersection of all complete assumption extensions. Asms+ = {α ∈ A | Asms attacks α } = {α ∈ A | in( Lab Asm) attacks α } = {α ∈ A | α ∈ out( Lab Asm)} = out( Lab Asm). A \ ( Asms ∪ Asms+ ) = {α ∈ A | α ∈ / in( Lab Asm), α ∈ / out( Lab Asm)} = {α ∈ A | α ∈ undec( Lab Asm)} = undec( Lab Asm). Proof of Proposition 37

• From right to left: Let Lab Asm be a grounded assumption labelling according to Definition 17. By Theorem 36 Asms = in( Lab Asm) is a grounded assumption extension of possibly non-flat ABA frameworks with Asms+ = out( Lab Asm) and A \ ( Asms ∪ Asms+ ) = undec( Lab Asm). By Theorem 6.2 in [4], for flat ABA frameworks Asms is a minimal (w.r.t. ⊆) complete assumption extension, and thus Asms is a grounded assumption extension as defined for flat ABA frameworks. By Theorem 5, Lab Asm is a grounded assumption labelling according to Definition 4. • From left to right: Let Lab Asm be a grounded assumption labelling according to Definition 4. By Theorem 5 Asms = in( Lab Asm) is a grounded assumption extension as defined for flat ABA frameworks with Asms+ = out( Lab Asm) and A \ ( Asms ∪ Asms+ ) = undec( Lab Asm), i.e. Asms is a minimal (w.r.t. ⊆) complete assumption extension. Let Asms be the intersection of all complete assumption extensions, i.e. a grounded assumption extension of possibly non-flat ABA frameworks. Since the grounded extension of a flat ABA framework is unique, Asms is unique and so is Asms . Thus,

148

C. Schulz, F. Toni / International Journal of Approximate Reasoning 84 (2017) 110–149

by Theorem 6.2 in [4] Asms = Asms. Then by Theorem 36 Lab Asm is a grounded assumption labelling according to Definition 17. Proof of Theorem 39 Analogous to the proof of Theorem 5, but using the definition of admissible assumption extensions and labellings of possibly non-flat ABA frameworks instead of complete assumption extensions and labellings of flat ABA frameworks, as well as Theorem 32 instead of Theorem 4. Proof of Theorem 41 1. By Theorem 5.5 in [4], Asms is a complete assumption extension. By Theorem 34, Lab Asm is a complete assumption labelling. Furthermore, since for all α ∈ A it holds that if α ∈ / Asms then Asms attacks α , it follows that Asms ∪ Asms+ = A. Then in( Lab Asm) ∪ out( Lab Asm) = A, so undec( Lab Asm) = ∅. Thus, Lab Asm is a stable assumption labelling. 2. By definition Lab Asm is a complete assumption labelling. By Theorem 34, Asms is a complete assumption extension. Since undec( Lab Asm) = ∅ it follows that for all α ∈ A, α ∈ in( Lab Asm) of α ∈ out( Lab Asm). And thus α ∈ Asms or α ∈ Asms+ . Thus, if α ∈/ Asms then Asms attacks α . Proof of Theorem 43 Analogous to the proof of Theorem 5, but using the definition of admissible assumption extensions and labellings of possibly non-flat ABA frameworks instead of complete assumption extensions of flat ABA frameworks, as well as the definition of preferred assumption extensions and labellings of possibly non-flat ABA frameworks instead of preferred assumption extensions and labellings of flat ABA frameworks, and Theorem 32 instead of Theorem 4. Proof of Theorem 45 Analogous to the proof of Theorem 5, but using the definition of admissible assumption extensions and labellings of possibly non-flat ABA frameworks instead of complete assumption extensions and labellings of flat ABA frameworks, and Theorem 32 instead of Theorem 4. Proof of Proposition 46 Since undec( Lab Asm) is minimal it follows that in( Lab Asm) ∪ out( Lab Asm) is maximal among all admissible assumption labellings. Assume by contradiction that there exists an admissible assumption labelling Lab Asm such that in( Lab Asm) ⊂ in( Lab Asm ). Then for all α ∈ A such that in( Lab Asm) attacks α , in( Lab Asm ) also attacks α . Thus, out( Lab Asm) ⊆ out( Lab Asm ). It follows that in( Lab Asm) ∪ out( Lab Asm) ⊂ in( Lab Asm ) ∪ out( Lab Asm ). Contradiction. References [1] T.J.M. Bench-Capon, P.E. Dunne, Argumentation in artificial intelligence, Artif. Intell. 171 (10–15) (2007) 619–641. [2] I. Rahwan, G.R. Simari, Argumentation in Artificial Intelligence, Springer US, 2009. [3] P.M. Dung, On the acceptability of arguments and its fundamental role in nonmonotonic reasoning, logic programming and n-person games, Artif. Intell. 77 (2) (1995) 321–357. [4] A. Bondarenko, P.M. Dung, R.A. Kowalski, F. Toni, An abstract, argumentation-theoretic approach to default reasoning, Artif. Intell. 93 (1–2) (1997) 63–101. [5] P. Besnard, A. Hunter, A logic-based theory of deductive arguments, Artif. Intell. 128 (1–2) (2001) 203–235. [6] A.J. García, G.R. Simari, Defeasible logic programming: an argumentative approach, Theory Pract. Log. Program. 4 (1–2) (2004) 95–138. [7] S. Modgil, H. Prakken, A general account of argumentation with preferences, Artif. Intell. 195 (2013) 361–397. [8] P. Besnard, A.J. García, A. Hunter, S. Modgil, H. Prakken, G.R. Simari, F. Toni, Introduction to structured argumentation, Argument Comput. 5 (1) (2014) 1–4. [9] P.M. Dung, R.A. Kowalski, F. Toni, Assumption-based argumentation, in: G.R. Simari, I. Rahwan (Eds.), Argumentation in Artificial Intelligence, Springer US, 2009, pp. 199–218. [10] F. Toni, A tutorial on assumption-based argumentation, Argument Comput. 5 (1) (2014) 89–117. [11] D. Gaertner, F. Toni, Preferences and assumption-based argumentation for conflict-free normative agents, in: Revised Selected and Invited Papers of the 4th International Workshop on Argumentation in Multi-Agent Systems, ArgMAS’07, 2007, pp. 94–113. [12] M. Morge, P. Mancarella, Assumption-based argumentation for the minimal concession strategy, in: Revised Selected and Invited Papers of the 6th International Workshop on Argumentation in Multi-Agent Systems, ArgMAS’09, 2009, pp. 114–133. [13] A. Hussain, F. Toni, Assumption-based argumentation for communicating agents, in: Papers from the 2009 AAAI Fall Symposium on the Uses of Computational Argumentation, 2009. [14] X. Fan, F. Toni, A. Hussain, Two-agent conflict resolution with assumption-based argumentation, in: Proceedings of the 3rd International Conference on Computational Models of Argument, COMMA’10, 2010, pp. 231–242. [15] M. Morge, P. Mancarella, Arguing over goals for negotiation: adopting an assumption-based argumentation decision support system, Group Decis. Negot. 23 (5) (2012) 979–1012. [16] X. Fan, F. Toni, A general framework for sound assumption-based argumentation dialogues, Artif. Intell. 216 (2014) 20–54.

C. Schulz, F. Toni / International Journal of Approximate Reasoning 84 (2017) 110–149

149

[17] X. Fan, F. Toni, Decision making with assumption-based argumentation, in: Revised Selected Papers of the 2nd International Workshop on Theory and Applications of Formal Argumentation, TAFA’13, 2013, pp. 127–142. [18] X. Fan, R. Craven, R. Singer, F. Toni, M. Williams, Assumption-based argumentation for decision-making with preferences: a medical case study, in: Proceedings of the 14th International Workshop on Computational Logic in Multi-Agent Systems, CLIMA’13, 2013, pp. 374–390. [19] X. Fan, F. Toni, A. Mocanu, M. Williams, Dialogical two-agent decision making with assumption-based argumentation, in: Proceedings of the 13th International Conference on Autonomous Agents and Multi-Agent Systems, AAMAS’14, 2014, pp. 533–540. [20] F. Toni, Reasoning on the web with assumption-based argumentation, in: Proceedings of the 8th International Summer School on Reasoning Web, 2012, pp. 370–386. [21] X. Fan, F. Toni, On computing explanations in argumentation, in: Proceedings of the 29th AAAI Conference on Artificial Intelligence, AAAI’15, 2015, pp. 1496–1502. [22] C. Schulz, F. Toni, Justifying answer sets using argumentation, Theory Pract. Log. Program. 16 (01) (2016) 59–110. [23] P.M. Dung, P. Mancarella, F. Toni, Computing ideal sceptical argumentation, Artif. Intell. 171 (10–15) (2007) 642–674. [24] M. Caminada, S. Sá, J. Alcântara, W. Dvoˇrák, On the difference between assumption-based argumentation and abstract argumentation, IfCoLog J. Log. Appl. 2 (1) (2015) 16–34. [25] M. Caminada, D.M. Gabbay, A logical account of formal argumentation, Stud. Log. 93 (2–3) (2009) 109–145. [26] P. Baroni, M. Caminada, M. Giacomin, An introduction to argumentation semantics, Knowl. Eng. Rev. 26 (04) (2011) 365–410. [27] C. Schulz, F. Toni, Complete assumption labellings, in: Proceedings of the 5th International Conference on Computational Models of Argument, COMMA’14, 2014, pp. 405–412. [28] C. Schulz, F. Toni, Logic programming in assumption-based argumentation revisited – semantics and graphical representation, in: Proceedings of the 29th AAAI Conference on Artificial Intelligence, AAAI’15, 2015, pp. 1569–1575. [29] M. Caminada, C. Schulz, On the equivalence between assumption-based argumentation and logic programming, in: Proceedings of the 1st International Workshop on Argumentation and Logic Programming, ArgLP’15, 2015. [30] M. Caminada, Semi-stable semantics, in: Proceedings of the 1st International Conference on Computational Models of Argument, COMMA’06, 2006, pp. 121–130. [31] M. Caminada, On the issue of reinstatement in argumentation, in: Proceedings of the 10th European Conference on Logics in Artificial Intelligence, JELIA’06, 2006, pp. 111–123. [32] M. Caminada, A labelling approach for ideal and stage semantics, Argument Comput. 2 (1) (2011) 1–21. [33] P.M. Dung, R.A. Kowalski, F. Toni, Dialectic proof procedures for assumption-based, admissible argumentation, Artif. Intell. 170 (2) (2006) 114–159. [34] F. Toni, A generalised framework for dispute derivations in assumption-based argumentation, Artif. Intell. 195 (2013) 1–43. [35] J. Heyninck, C. Straßer, Relations between assumption-based approaches in nonmonotonic logic and formal argumentation, in: Proceedings of the 16th International Workshop on Non-Monotonic Reasoning, NMR’16, 2016. [36] R. Craven, F. Toni, Argument graphs and assumption-based argumentation, Artif. Intell. 233 (2016) 1–59. [37] P. Baroni, M. Giacomin, On principle-based evaluation of extension-based argumentation semantics, Artif. Intell. 171 (10–15) (2007) 675–700. [38] M. Caminada, Strong admissibility revisited, in: Proceedings of the 5th International Conference on Computational Models of Argument, COMMA’14, 2014, pp. 197–208. [39] R.C. Moore, Semantical considerations on nonmonotonic logic, Artif. Intell. 25 (1) (1985) 75–94. [40] M. Caminada, S. Sá, J. Alcântara, W. Dvoˇrák, On the equivalence between logic programming semantics and argumentation semantics, Int. J. Approx. Reason. 58 (2015) 87–111. [41] T.C. Przymusinski, Stable semantics for disjunctive programs, New Gener. Comput. 9 (3/4) (1991) 401–424. [42] A.V. Gelder, K.A. Ross, J.S. Schlipf, The well-founded semantics for general logic programs, J. ACM 38 (3) (1991) 620–650. [43] J.-H. You, L.Y. Yuan, Three-valued formalization of logic programming: is it needed?, in: Proceedings of the 9th ACM SIGACT–SIGMOD–SIGART Symposium on Principles of Database Systems, PODS’90, 1990, pp. 172–182. [44] C. Schulz, K. Satoh, F. Toni, Characterising and explaining inconsistency in logic programs, in: Proceedings of the 13th International Conference on Logic Programming and Nonmonotonic Reasoning, LPNMR’15, 2015, pp. 467–479. [45] Y. Dimopoulos, A. Torres, Graph theoretical structures in logic programs and default theories, Theor. Comput. Sci. 170 (1–2) (1996) 209–244. [46] S. Costantini, Comparing different graph representations of logic programs under the answer set semantics, in: Proceedings of the 1st International Workshop on Answer Set Programming: Towards Efficient and Scalable Knowledge Representation and Reasoning, ASP’01, 2001. [47] T. Linke, V. Sarsakov, Suitable graphs for answer set programming, in: Proceedings of the 11th International Conference on Logic for Programming, Artificial Intelligence, and Reasoning, LPAR’04, 2004, pp. 154–168. [48] K. Konczak, T. Linke, T. Schaub, Graphs and colorings for answer set programming, Theory Pract. Log. Program. 6 (1–2) (2006) 61–106. [49] T. Wakaki, Assumption-based argumentation equipped with preferences, in: Proceedings of the 17th International Conference on Principles and Practice of Multi-Agent Systems, PRIMA’14, 2014, pp. 116–132. [50] K. Cyras, F. Toni, ABA+: assumption-based argumentation with preferences, in: Proceedings of the 15th International Conference on Principles of Knowledge Representation and Reasoning, KR’16, 2016, pp. 553–556. [51] S. Modgil, M. Caminada, Proof theories and algorithms for abstract argumentation frameworks, in: G. Simari, I. Rahwan (Eds.), Argumentation in Artificial Intelligence, Springer US, 2009, pp. 105–129. [52] T. Wakaki, K. Nitta, Computing argumentation semantics in answer set programming, in: Revised Selected Papers of the 22nd Annual Conference of the Japanese Society for Artificial Intelligence, JSAI’08, 2008, pp. 254–269. [53] G. Charwat, W. Dvoˇrák, S.A. Gaggl, J.P. Wallner, S. Woltran, Methods for solving reasoning problems in abstract argumentation – a survey, Artif. Intell. 220 (2015) 28–63.